git.oblomov.eu Git - wine/blob - dlls/wined3d/arb_program_shader.c

   1 /*
   2  * Pixel and vertex shaders implementation using ARB_vertex_program
   3  * and ARB_fragment_program GL extensions.
   4  *
   5  * Copyright 2002-2003 Jason Edmeades
   6  * Copyright 2002-2003 Raphael Junqueira
   7  * Copyright 2004 Christian Costa
   8  * Copyright 2005 Oliver Stieber
   9  * Copyright 2006 Ivan Gyurdiev
  10  * Copyright 2006 Jason Green
  11  * Copyright 2006 Henri Verbeet
  12  * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
  13  *
  14  * This library is free software; you can redistribute it and/or
  15  * modify it under the terms of the GNU Lesser General Public
  16  * License as published by the Free Software Foundation; either
  17  * version 2.1 of the License, or (at your option) any later version.
  18  *
  19  * This library is distributed in the hope that it will be useful,
  20  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  21  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  * Lesser General Public License for more details.
  23  *
  24  * You should have received a copy of the GNU Lesser General Public
  25  * License along with this library; if not, write to the Free Software
  26  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  27  */
  28
  29 #include "config.h"
  30
  31 #include <math.h>
  32 #include <stdio.h>
  33
  34 #include "wined3d_private.h"
  35
  36 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  37 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
  38 WINE_DECLARE_DEBUG_CHANNEL(d3d_caps);
  39 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  40
  41 #define GLINFO_LOCATION      (*gl_info)
  42
  43 /* We have to subtract any other PARAMs that we might use in our shader programs.
  44  * ATI seems to count 2 implicit PARAMs when we use fog and NVIDIA counts 1,
  45  * and we reference one row of the PROJECTION matrix which counts as 1 PARAM. */
  46 #define ARB_SHADER_RESERVED_VS_CONSTS 3
  47
  48 /* The arb shader only loads the bump mapping environment matrix into the shader if it finds
  49  * a free constant to do that, so only reduce the number of available constants by 2 for the fog states.
  50  */
  51 #define ARB_SHADER_RESERVED_PS_CONSTS 2
  52
  53 /* Internally used shader constants. Applications can use constants 0 to GL_LIMITS(vshader_constantsF) - 1,
  54  * so upload them above that
  55  */
  56 #define ARB_SHADER_PRIVCONST_BASE (GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS)
  57 #define ARB_SHADER_PRIVCONST_POS ARB_SHADER_PRIVCONST_BASE + 0
  58
  59 /* ARB_program_shader private data */
  60 struct shader_arb_priv {
  61     GLuint                  current_vprogram_id;
  62     GLuint                  current_fprogram_id;
  63     GLuint                  depth_blt_vprogram_id;
  64     GLuint                  depth_blt_fprogram_id[tex_type_count];
  65     BOOL                    use_arbfp_fixed_func;
  66     struct hash_table_t     *fragment_shaders;
  67 };
  68
  69 /********************************************************
  70  * ARB_[vertex/fragment]_program helper functions follow
  71  ********************************************************/
  72
  73 /**
  74  * Loads floating point constants into the currently set ARB_vertex/fragment_program.
  75  * When constant_list == NULL, it will load all the constants.
  76  *
  77  * @target_type should be either GL_VERTEX_PROGRAM_ARB (for vertex shaders)
  78  *  or GL_FRAGMENT_PROGRAM_ARB (for pixel shaders)
  79  */
  80 static unsigned int shader_arb_load_constantsF(IWineD3DBaseShaderImpl* This, const WineD3D_GL_Info *gl_info,
  81         GLuint target_type, unsigned int max_constants, const float *constants, char *dirty_consts)
  82 {
  83     DWORD shader_version = This->baseShader.reg_maps.shader_version;
  84     local_constant* lconst;
  85     DWORD i, j;
  86     unsigned int ret;
  87
  88     if (TRACE_ON(d3d_shader)) {
  89         for(i = 0; i < max_constants; i++) {
  90             if(!dirty_consts[i]) continue;
  91             TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
  92                         constants[i * 4 + 0], constants[i * 4 + 1],
  93                         constants[i * 4 + 2], constants[i * 4 + 3]);
  94         }
  95     }
  96     /* In 1.X pixel shaders constants are implicitly clamped in the range [-1;1] */
  97     if (target_type == GL_FRAGMENT_PROGRAM_ARB && WINED3DSHADER_VERSION_MAJOR(shader_version) == 1)
  98     {
  99         float lcl_const[4];
 100         for(i = 0; i < max_constants; i++) {
 101             if(!dirty_consts[i]) continue;
 102             dirty_consts[i] = 0;
 103
 104             j = 4 * i;
 105             if(constants[j + 0] > 1.0) lcl_const[0] = 1.0;
 106             else if(constants[j + 0] < -1.0) lcl_const[0] = -1.0;
 107             else lcl_const[0] = constants[j + 0];
 108
 109             if(constants[j + 1] > 1.0) lcl_const[1] = 1.0;
 110             else if(constants[j + 1] < -1.0) lcl_const[1] = -1.0;
 111             else lcl_const[1] = constants[j + 1];
 112
 113             if(constants[j + 2] > 1.0) lcl_const[2] = 1.0;
 114             else if(constants[j + 2] < -1.0) lcl_const[2] = -1.0;
 115             else lcl_const[2] = constants[j + 2];
 116
 117             if(constants[j + 3] > 1.0) lcl_const[3] = 1.0;
 118             else if(constants[j + 3] < -1.0) lcl_const[3] = -1.0;
 119             else lcl_const[3] = constants[j + 3];
 120
 121             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, lcl_const));
 122         }
 123     } else {
 124         if(GL_SUPPORT(EXT_GPU_PROGRAM_PARAMETERS)) {
 125             /* TODO: Benchmark if we're better of with finding the dirty constants ourselves,
 126              * or just reloading *all* constants at once
 127              *
 128             GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, 0, max_constants, constants));
 129              */
 130             for(i = 0; i < max_constants; i++) {
 131                 if(!dirty_consts[i]) continue;
 132
 133                 /* Find the next block of dirty constants */
 134                 dirty_consts[i] = 0;
 135                 j = i;
 136                 for(i++; (i < max_constants) && dirty_consts[i]; i++) {
 137                     dirty_consts[i] = 0;
 138                 }
 139
 140                 GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, j, i - j, constants + (j * 4)));
 141             }
 142         } else {
 143             for(i = 0; i < max_constants; i++) {
 144                 if(dirty_consts[i]) {
 145                     dirty_consts[i] = 0;
 146                     GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
 147                 }
 148             }
 149         }
 150     }
 151     checkGLcall("glProgramEnvParameter4fvARB()");
 152
 153     /* Load immediate constants */
 154     if(This->baseShader.load_local_constsF) {
 155         if (TRACE_ON(d3d_shader)) {
 156             LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 157                 GLfloat* values = (GLfloat*)lconst->value;
 158                 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
 159                         values[0], values[1], values[2], values[3]);
 160             }
 161         }
 162         /* Immediate constants are clamped for 1.X shaders at loading times */
 163         ret = 0;
 164         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 165             dirty_consts[lconst->idx] = 1; /* Dirtify so the non-immediate constant overwrites it next time */
 166             ret = max(ret, lconst->idx + 1);
 167             GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, lconst->idx, (GLfloat*)lconst->value));
 168         }
 169         checkGLcall("glProgramEnvParameter4fvARB()");
 170         return ret; /* The loaded immediate constants need reloading for the next shader */
 171     } else {
 172         return 0; /* No constants are dirty now */
 173     }
 174 }
 175
 176 /**
 177  * Loads the texture dimensions for NP2 fixup into the currently set ARB_[vertex/fragment]_programs.
 178  */
 179 static void shader_arb_load_np2fixup_constants(
 180     IWineD3DDevice* device,
 181     char usePixelShader,
 182     char useVertexShader) {
 183     /* not implemented */
 184 }
 185
 186 /**
 187  * Loads the app-supplied constants into the currently set ARB_[vertex/fragment]_programs.
 188  *
 189  * We only support float constants in ARB at the moment, so don't
 190  * worry about the Integers or Booleans
 191  */
 192 static void shader_arb_load_constants(
 193     IWineD3DDevice* device,
 194     char usePixelShader,
 195     char useVertexShader) {
 196
 197     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
 198     IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
 199     const WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
 200     unsigned char i;
 201
 202     if (useVertexShader) {
 203         IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
 204
 205         /* Load DirectX 9 float constants for vertex shader */
 206         deviceImpl->highest_dirty_vs_const = shader_arb_load_constantsF(
 207                 vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
 208                 deviceImpl->highest_dirty_vs_const,
 209                 stateBlock->vertexShaderConstantF,
 210                 deviceImpl->activeContext->vshader_const_dirty);
 211
 212         /* Upload the position fixup */
 213         GL_EXTCALL(glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, ARB_SHADER_PRIVCONST_POS, deviceImpl->posFixup));
 214     }
 215
 216     if (usePixelShader) {
 217
 218         IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
 219         IWineD3DPixelShaderImpl *psi = (IWineD3DPixelShaderImpl *) pshader;
 220
 221         /* Load DirectX 9 float constants for pixel shader */
 222         deviceImpl->highest_dirty_ps_const = shader_arb_load_constantsF(
 223                 pshader, gl_info, GL_FRAGMENT_PROGRAM_ARB,
 224                 deviceImpl->highest_dirty_ps_const,
 225                 stateBlock->pixelShaderConstantF,
 226                 deviceImpl->activeContext->pshader_const_dirty);
 227
 228         for(i = 0; i < psi->numbumpenvmatconsts; i++) {
 229             /* The state manager takes care that this function is always called if the bump env matrix changes
 230              */
 231             const float *data = (const float *)&stateBlock->textureState[(int) psi->bumpenvmatconst[i].texunit][WINED3DTSS_BUMPENVMAT00];
 232             GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->bumpenvmatconst[i].const_num, data));
 233             deviceImpl->activeContext->pshader_const_dirty[psi->bumpenvmatconst[i].const_num] = 1;
 234
 235             if (psi->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED)
 236             {
 237                 /* WINED3DTSS_BUMPENVLSCALE and WINED3DTSS_BUMPENVLOFFSET are next to each other.
 238                  * point gl to the scale, and load 4 floats. x = scale, y = offset, z and w are junk, we
 239                  * don't care about them. The pointers are valid for sure because the stateblock is bigger.
 240                  * (they're WINED3DTSS_TEXTURETRANSFORMFLAGS and WINED3DTSS_ADDRESSW, so most likely 0 or NaN
 241                  */
 242                 const float *scale = (const float *)&stateBlock->textureState[(int) psi->luminanceconst[i].texunit][WINED3DTSS_BUMPENVLSCALE];
 243                 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->luminanceconst[i].const_num, scale));
 244                 deviceImpl->activeContext->pshader_const_dirty[psi->luminanceconst[i].const_num] = 1;
 245             }
 246         }
 247     }
 248 }
 249
 250 static void shader_arb_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count)
 251 {
 252     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 253
 254     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 255      * context. On a context switch the old context will be fully dirtified */
 256     memset(This->activeContext->vshader_const_dirty + start, 1,
 257             sizeof(*This->activeContext->vshader_const_dirty) * count);
 258     This->highest_dirty_vs_const = max(This->highest_dirty_vs_const, start + count + 1);
 259 }
 260
 261 static void shader_arb_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count)
 262 {
 263     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 264
 265     /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
 266      * context. On a context switch the old context will be fully dirtified */
 267     memset(This->activeContext->pshader_const_dirty + start, 1,
 268             sizeof(*This->activeContext->pshader_const_dirty) * count);
 269     This->highest_dirty_ps_const = max(This->highest_dirty_ps_const, start + count + 1);
 270 }
 271
 272 /* Generate the variable & register declarations for the ARB_vertex_program output target */
 273 static void shader_generate_arb_declarations(IWineD3DBaseShader *iface, const shader_reg_maps *reg_maps,
 274         SHADER_BUFFER *buffer, const WineD3D_GL_Info *gl_info)
 275 {
 276     IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
 277     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 278     DWORD i, cur;
 279     char pshader = shader_is_pshader_version(reg_maps->shader_version);
 280     unsigned max_constantsF = min(This->baseShader.limits.constant_float,
 281             (pshader ? GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS :
 282                        GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS));
 283     UINT extra_constants_needed = 0;
 284     const local_constant *lconst;
 285
 286     /* Temporary Output register */
 287     shader_addline(buffer, "TEMP TMP_OUT;\n");
 288
 289     for(i = 0; i < This->baseShader.limits.temporary; i++) {
 290         if (reg_maps->temporary[i])
 291             shader_addline(buffer, "TEMP R%u;\n", i);
 292     }
 293
 294     for (i = 0; i < This->baseShader.limits.address; i++) {
 295         if (reg_maps->address[i])
 296             shader_addline(buffer, "ADDRESS A%d;\n", i);
 297     }
 298
 299     for(i = 0; i < This->baseShader.limits.texcoord; i++) {
 300         if (reg_maps->texcoord[i])
 301             shader_addline(buffer,"TEMP T%u;\n", i);
 302     }
 303
 304     /* Texture coordinate registers must be pre-loaded */
 305     for (i = 0; i < This->baseShader.limits.texcoord; i++) {
 306         if (reg_maps->texcoord[i])
 307             shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
 308     }
 309
 310     for(i = 0; i < (sizeof(reg_maps->bumpmat) / sizeof(reg_maps->bumpmat[0])); i++) {
 311         IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
 312         if(!reg_maps->bumpmat[i]) continue;
 313
 314         cur = ps->numbumpenvmatconsts;
 315         ps->bumpenvmatconst[cur].const_num = -1;
 316         ps->bumpenvmatconst[cur].texunit = i;
 317         ps->luminanceconst[cur].const_num = -1;
 318         ps->luminanceconst[cur].texunit = i;
 319
 320         /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
 321          * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
 322          * bump mapping.
 323          */
 324         if(max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
 325             ps->bumpenvmatconst[cur].const_num = max_constantsF + extra_constants_needed;
 326             shader_addline(buffer, "PARAM bumpenvmat%d = program.env[%d];\n",
 327                            i, ps->bumpenvmatconst[cur].const_num);
 328             extra_constants_needed++;
 329
 330             if(reg_maps->luminanceparams && max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
 331                 ((IWineD3DPixelShaderImpl *)This)->luminanceconst[cur].const_num = max_constantsF + extra_constants_needed;
 332                 shader_addline(buffer, "PARAM luminance%d = program.env[%d];\n",
 333                                i, ps->luminanceconst[cur].const_num);
 334                 extra_constants_needed++;
 335             } else if(reg_maps->luminanceparams) {
 336                 FIXME("No free constant to load the luminance parameters\n");
 337             }
 338         } else {
 339             FIXME("No free constant found to load environment bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
 340         }
 341
 342         ps->numbumpenvmatconsts = cur + 1;
 343     }
 344
 345     if(device->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] && pshader) {
 346         shader_addline(buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
 347                         srgb_mul_low, srgb_mul_low, srgb_mul_low);
 348         shader_addline(buffer, "PARAM srgb_comparison =  {%f, %f, %f, %f};\n",
 349                         srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
 350         shader_addline(buffer, "PARAM srgb_pow =  {%f, %f, %f, 1.0};\n",
 351                        srgb_pow, srgb_pow, srgb_pow);
 352         shader_addline(buffer, "PARAM srgb_mul_hi =  {%f, %f, %f, 1.0};\n",
 353                        srgb_mul_high, srgb_mul_high, srgb_mul_high);
 354         shader_addline(buffer, "PARAM srgb_sub_hi =  {%f, %f, %f, 0.0};\n",
 355                        srgb_sub_high, srgb_sub_high, srgb_sub_high);
 356     }
 357
 358     /* Load local constants using the program-local space,
 359      * this avoids reloading them each time the shader is used
 360      */
 361     if(!This->baseShader.load_local_constsF) {
 362         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
 363             shader_addline(buffer, "PARAM C%u = program.local[%u];\n", lconst->idx,
 364                            lconst->idx);
 365         }
 366     }
 367
 368     /* we use the array-based constants array if the local constants are marked for loading,
 369      * because then we use indirect addressing, or when the local constant list is empty,
 370      * because then we don't know if we're using indirect addressing or not. If we're hardcoding
 371      * local constants do not declare the loaded constants as an array because ARB compilers usually
 372      * do not optimize unused constants away
 373      */
 374     if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
 375         /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
 376         shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
 377                     max_constantsF, max_constantsF - 1);
 378     } else {
 379         for(i = 0; i < max_constantsF; i++) {
 380             if(!shader_constant_is_local(This, i)) {
 381                 shader_addline(buffer, "PARAM C%d = program.env[%d];\n",i, i);
 382             }
 383         }
 384     }
 385 }
 386
 387 static const char * const shift_tab[] = {
 388     "dummy",     /*  0 (none) */
 389     "coefmul.x", /*  1 (x2)   */
 390     "coefmul.y", /*  2 (x4)   */
 391     "coefmul.z", /*  3 (x8)   */
 392     "coefmul.w", /*  4 (x16)  */
 393     "dummy",     /*  5 (x32)  */
 394     "dummy",     /*  6 (x64)  */
 395     "dummy",     /*  7 (x128) */
 396     "dummy",     /*  8 (d256) */
 397     "dummy",     /*  9 (d128) */
 398     "dummy",     /* 10 (d64)  */
 399     "dummy",     /* 11 (d32)  */
 400     "coefdiv.w", /* 12 (d16)  */
 401     "coefdiv.z", /* 13 (d8)   */
 402     "coefdiv.y", /* 14 (d4)   */
 403     "coefdiv.x"  /* 15 (d2)   */
 404 };
 405
 406 static void shader_arb_get_write_mask(const struct wined3d_shader_instruction *ins,
 407         const struct wined3d_shader_dst_param *dst, char *write_mask)
 408 {
 409     char *ptr = write_mask;
 410     char vshader = shader_is_vshader_version(ins->reg_maps->shader_version);
 411
 412     if (vshader && dst->register_type == WINED3DSPR_ADDR)
 413     {
 414         *ptr++ = '.';
 415         *ptr++ = 'x';
 416     }
 417     else if (dst->write_mask != WINED3DSP_WRITEMASK_ALL)
 418     {
 419         *ptr++ = '.';
 420         if (dst->write_mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 421         if (dst->write_mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 422         if (dst->write_mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 423         if (dst->write_mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 424     }
 425
 426     *ptr = '\0';
 427 }
 428
 429 static void shader_arb_get_swizzle(const DWORD param, BOOL fixup, char *swizzle_str) {
 430     /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
 431      * but addressed as "rgba". To fix this we need to swap the register's x
 432      * and z components. */
 433     const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
 434     char *ptr = swizzle_str;
 435
 436     /* swizzle bits fields: wwzzyyxx */
 437     DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
 438     DWORD swizzle_x = swizzle & 0x03;
 439     DWORD swizzle_y = (swizzle >> 2) & 0x03;
 440     DWORD swizzle_z = (swizzle >> 4) & 0x03;
 441     DWORD swizzle_w = (swizzle >> 6) & 0x03;
 442
 443     /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
 444      * generate a swizzle string. Unless we need to our own swizzling. */
 445     if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) != swizzle || fixup) {
 446         *ptr++ = '.';
 447         if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
 448             *ptr++ = swizzle_chars[swizzle_x];
 449         } else {
 450             *ptr++ = swizzle_chars[swizzle_x];
 451             *ptr++ = swizzle_chars[swizzle_y];
 452             *ptr++ = swizzle_chars[swizzle_z];
 453             *ptr++ = swizzle_chars[swizzle_w];
 454         }
 455     }
 456
 457     *ptr = '\0';
 458 }
 459
 460 static void shader_arb_get_register_name(IWineD3DBaseShader *iface, WINED3DSHADER_PARAM_REGISTER_TYPE register_type,
 461         UINT register_idx, BOOL rel_addr, char *register_name, BOOL *is_color)
 462 {
 463     /* oPos, oFog and oPts in D3D */
 464     static const char * const rastout_reg_names[] = {"TMP_OUT", "result.fogcoord", "result.pointsize"};
 465     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
 466     DWORD shader_version = This->baseShader.reg_maps.shader_version;
 467     BOOL pshader = shader_is_pshader_version(shader_version);
 468
 469     *is_color = FALSE;
 470
 471     switch (register_type)
 472     {
 473         case WINED3DSPR_TEMP:
 474             sprintf(register_name, "R%u", register_idx);
 475             break;
 476
 477         case WINED3DSPR_INPUT:
 478             if (pshader)
 479             {
 480                 if (register_idx == 0) strcpy(register_name, "fragment.color.primary");
 481                 else strcpy(register_name, "fragment.color.secondary");
 482             }
 483             else
 484             {
 485                 if (((IWineD3DVertexShaderImpl *)This)->cur_args->swizzle_map & (1 << register_idx)) *is_color = TRUE;
 486                 sprintf(register_name, "vertex.attrib[%u]", register_idx);
 487             }
 488             break;
 489
 490         case WINED3DSPR_CONST:
 491             if (!pshader && rel_addr)
 492             {
 493                 UINT rel_offset = ((IWineD3DVertexShaderImpl *)This)->rel_offset;
 494                 if (register_idx >= rel_offset)
 495                     sprintf(register_name, "C[A0.x + %u]", register_idx - rel_offset);
 496                 else
 497                     sprintf(register_name, "C[A0.x - %u]", -register_idx + rel_offset);
 498             }
 499             else
 500             {
 501                 if (This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF))
 502                     sprintf(register_name, "C[%u]", register_idx);
 503                 else
 504                     sprintf(register_name, "C%u", register_idx);
 505             }
 506             break;
 507
 508         case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
 509             if (pshader) sprintf(register_name, "T%u", register_idx);
 510             else  sprintf(register_name, "A%u", register_idx);
 511             break;
 512
 513         case WINED3DSPR_COLOROUT:
 514             if (register_idx == 0)
 515             {
 516                 strcpy(register_name, "TMP_COLOR");
 517             }
 518             else
 519             {
 520                 /* TODO: See GL_ARB_draw_buffers */
 521                 FIXME("Unsupported write to render target %u\n", register_idx);
 522                 sprintf(register_name, "unsupported_register");
 523             }
 524             break;
 525
 526         case WINED3DSPR_RASTOUT:
 527             sprintf(register_name, "%s", rastout_reg_names[register_idx]);
 528             break;
 529
 530         case WINED3DSPR_DEPTHOUT:
 531             strcpy(register_name, "result.depth");
 532             break;
 533
 534         case WINED3DSPR_ATTROUT:
 535             if (pshader) sprintf(register_name, "oD[%u]", register_idx);
 536             else if (register_idx == 0) strcpy(register_name, "result.color.primary");
 537             else strcpy(register_name, "result.color.secondary");
 538             break;
 539
 540         case WINED3DSPR_TEXCRDOUT:
 541             if (pshader) sprintf(register_name, "oT[%u]", register_idx);
 542             else sprintf(register_name, "result.texcoord[%u]", register_idx);
 543             break;
 544
 545         default:
 546             FIXME("Unhandled register type %#x[%u]\n", register_type, register_idx);
 547             sprintf(register_name, "unrecognized_register[%u]", register_idx);
 548             break;
 549     }
 550 }
 551
 552 static void shader_arb_add_src_param(const struct wined3d_shader_instruction *ins,
 553         DWORD param, char *str)
 554 {
 555     char register_name[255];
 556     char swizzle[6];
 557     BOOL is_color;
 558
 559     if ((param & WINED3DSP_SRCMOD_MASK) == WINED3DSPSM_NEG) strcat(str, " -");
 560     else strcat(str, " ");
 561
 562     shader_arb_get_register_name(ins->shader, shader_get_regtype(param), param & WINED3DSP_REGNUM_MASK,
 563             param & WINED3DSHADER_ADDRMODE_RELATIVE, register_name, &is_color);
 564     strcat(str, register_name);
 565
 566     shader_arb_get_swizzle(param, is_color, swizzle);
 567     strcat(str, swizzle);
 568 }
 569
 570 static void shader_arb_add_dst_param(const struct wined3d_shader_instruction *ins,
 571         const struct wined3d_shader_dst_param *wined3d_dst, char *str)
 572 {
 573     char register_name[255];
 574     char write_mask[6];
 575     BOOL is_color;
 576
 577     strcat(str, " ");
 578
 579     shader_arb_get_register_name(ins->shader, wined3d_dst->register_type,
 580             wined3d_dst->register_idx, wined3d_dst->has_rel_addr, register_name, &is_color);
 581     strcat(str, register_name);
 582
 583     shader_arb_get_write_mask(ins, wined3d_dst, write_mask);
 584     strcat(str, write_mask);
 585 }
 586
 587 static const char *shader_arb_get_fixup_swizzle(enum fixup_channel_source channel_source)
 588 {
 589     switch(channel_source)
 590     {
 591         case CHANNEL_SOURCE_ZERO: return "0";
 592         case CHANNEL_SOURCE_ONE: return "1";
 593         case CHANNEL_SOURCE_X: return "x";
 594         case CHANNEL_SOURCE_Y: return "y";
 595         case CHANNEL_SOURCE_Z: return "z";
 596         case CHANNEL_SOURCE_W: return "w";
 597         default:
 598             FIXME("Unhandled channel source %#x\n", channel_source);
 599             return "undefined";
 600     }
 601 }
 602
 603 static void gen_color_correction(SHADER_BUFFER *buffer, const char *reg, DWORD dst_mask,
 604                                  const char *one, const char *two, struct color_fixup_desc fixup)
 605 {
 606     DWORD mask;
 607
 608     if (is_yuv_fixup(fixup))
 609     {
 610         enum yuv_fixup yuv_fixup = get_yuv_fixup(fixup);
 611         FIXME("YUV fixup (%#x) not supported\n", yuv_fixup);
 612         return;
 613     }
 614
 615     mask = 0;
 616     if (fixup.x_source != CHANNEL_SOURCE_X) mask |= WINED3DSP_WRITEMASK_0;
 617     if (fixup.y_source != CHANNEL_SOURCE_Y) mask |= WINED3DSP_WRITEMASK_1;
 618     if (fixup.z_source != CHANNEL_SOURCE_Z) mask |= WINED3DSP_WRITEMASK_2;
 619     if (fixup.w_source != CHANNEL_SOURCE_W) mask |= WINED3DSP_WRITEMASK_3;
 620     mask &= dst_mask;
 621
 622     if (mask)
 623     {
 624         shader_addline(buffer, "SWZ %s, %s, %s, %s, %s, %s;\n", reg, reg,
 625                 shader_arb_get_fixup_swizzle(fixup.x_source), shader_arb_get_fixup_swizzle(fixup.y_source),
 626                 shader_arb_get_fixup_swizzle(fixup.z_source), shader_arb_get_fixup_swizzle(fixup.w_source));
 627     }
 628
 629     mask = 0;
 630     if (fixup.x_sign_fixup) mask |= WINED3DSP_WRITEMASK_0;
 631     if (fixup.y_sign_fixup) mask |= WINED3DSP_WRITEMASK_1;
 632     if (fixup.z_sign_fixup) mask |= WINED3DSP_WRITEMASK_2;
 633     if (fixup.w_sign_fixup) mask |= WINED3DSP_WRITEMASK_3;
 634     mask &= dst_mask;
 635
 636     if (mask)
 637     {
 638         char reg_mask[6];
 639         char *ptr = reg_mask;
 640
 641         if (mask != WINED3DSP_WRITEMASK_ALL)
 642         {
 643             *ptr++ = '.';
 644             if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
 645             if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
 646             if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
 647             if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
 648         }
 649         *ptr = '\0';
 650
 651         shader_addline(buffer, "MAD %s%s, %s, %s, -%s;\n", reg, reg_mask, reg, two, one);
 652     }
 653 }
 654
 655 static void shader_hw_sample(const struct wined3d_shader_instruction *ins, DWORD sampler_idx,
 656         const char *dst_str, const char *coord_reg, BOOL projected, BOOL bias)
 657 {
 658     SHADER_BUFFER *buffer = ins->buffer;
 659     DWORD sampler_type = ins->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
 660     const char *tex_type;
 661     IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)ins->shader;
 662     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 663
 664     switch(sampler_type) {
 665         case WINED3DSTT_1D:
 666             tex_type = "1D";
 667             break;
 668
 669         case WINED3DSTT_2D:
 670             if(device->stateBlock->textures[sampler_idx] &&
 671                IWineD3DBaseTexture_GetTextureDimensions(device->stateBlock->textures[sampler_idx]) == GL_TEXTURE_RECTANGLE_ARB) {
 672                 tex_type = "RECT";
 673             } else {
 674                 tex_type = "2D";
 675             }
 676             if(shader_is_pshader_version(ins->reg_maps->shader_version)) {
 677                const IWineD3DPixelShaderImpl* const ps = (const IWineD3DPixelShaderImpl*)This;
 678                if(ps->cur_args->np2_fixup & (1 << sampler_idx)) {
 679                    FIXME("NP2 texcoord fixup is currently not implemented in ARB mode (use GLSL instead).\n");
 680                }
 681             }
 682             break;
 683
 684         case WINED3DSTT_VOLUME:
 685             tex_type = "3D";
 686             break;
 687
 688         case WINED3DSTT_CUBE:
 689             tex_type = "CUBE";
 690             break;
 691
 692         default:
 693             ERR("Unexpected texture type %d\n", sampler_type);
 694             tex_type = "";
 695     }
 696
 697     if (bias) {
 698         /* Shouldn't be possible, but let's check for it */
 699         if(projected) FIXME("Biased and Projected texture sampling\n");
 700         /* TXB takes the 4th component of the source vector automatically, as d3d. Nothing more to do */
 701         shader_addline(buffer, "TXB %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 702     } else if (projected) {
 703         shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 704     } else {
 705         shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
 706     }
 707
 708     if (shader_is_pshader_version(ins->reg_maps->shader_version))
 709     {
 710         IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)ins->shader;
 711         gen_color_correction(buffer, dst_str, ins->dst[0].write_mask,
 712                 "one", "coefmul.x", ps->cur_args->color_fixup[sampler_idx]);
 713     }
 714 }
 715
 716 static void pshader_gen_input_modifier_line (
 717     IWineD3DBaseShader *iface,
 718     SHADER_BUFFER* buffer,
 719     const DWORD instr,
 720     int tmpreg,
 721     char *outregstr) {
 722
 723     /* Generate a line that does the input modifier computation and return the input register to use */
 724     BOOL is_color = FALSE;
 725     char regstr[256];
 726     char swzstr[20];
 727     int insert_line;
 728
 729     /* Assume a new line will be added */
 730     insert_line = 1;
 731
 732     /* Get register name */
 733     shader_arb_get_register_name(iface, shader_get_regtype(instr), instr & WINED3DSP_REGNUM_MASK,
 734             instr & WINED3DSHADER_ADDRMODE_RELATIVE, regstr, &is_color);
 735     shader_arb_get_swizzle(instr, is_color, swzstr);
 736
 737     switch (instr & WINED3DSP_SRCMOD_MASK) {
 738     case WINED3DSPSM_NONE:
 739         sprintf(outregstr, "%s%s", regstr, swzstr);
 740         insert_line = 0;
 741         break;
 742     case WINED3DSPSM_NEG:
 743         sprintf(outregstr, "-%s%s", regstr, swzstr);
 744         insert_line = 0;
 745         break;
 746     case WINED3DSPSM_BIAS:
 747         shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
 748         break;
 749     case WINED3DSPSM_BIASNEG:
 750         shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
 751         break;
 752     case WINED3DSPSM_SIGN:
 753         shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
 754         break;
 755     case WINED3DSPSM_SIGNNEG:
 756         shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
 757         break;
 758     case WINED3DSPSM_COMP:
 759         shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
 760         break;
 761     case WINED3DSPSM_X2:
 762         shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
 763         break;
 764     case WINED3DSPSM_X2NEG:
 765         shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
 766         break;
 767     case WINED3DSPSM_DZ:
 768         shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
 769         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 770         break;
 771     case WINED3DSPSM_DW:
 772         shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
 773         shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
 774         break;
 775     default:
 776         sprintf(outregstr, "%s%s", regstr, swzstr);
 777         insert_line = 0;
 778     }
 779
 780     /* Return modified or original register, with swizzle */
 781     if (insert_line)
 782         sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
 783 }
 784
 785 static inline void pshader_gen_output_modifier_line(SHADER_BUFFER *buffer, int saturate, const char *write_mask,
 786         int shift, const char *regstr)
 787 {
 788     /* Generate a line that does the output modifier computation */
 789     shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
 790         regstr, write_mask, regstr, shift_tab[shift]);
 791 }
 792
 793 static void pshader_hw_bem(const struct wined3d_shader_instruction *ins)
 794 {
 795     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
 796     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 797     SHADER_BUFFER *buffer = ins->buffer;
 798     char dst_name[50];
 799     char src_name[2][50];
 800     char dst_wmask[20];
 801     DWORD sampler_code = dst->register_idx;
 802     BOOL has_bumpmat = FALSE;
 803     BOOL is_color;
 804     int i;
 805
 806     for(i = 0; i < This->numbumpenvmatconsts; i++) {
 807         if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
 808                 && This->bumpenvmatconst[i].texunit == sampler_code)
 809         {
 810             has_bumpmat = TRUE;
 811             break;
 812         }
 813     }
 814
 815     shader_arb_get_register_name(ins->shader, dst->register_type,
 816             dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
 817     shader_arb_get_write_mask(ins, dst, dst_wmask);
 818     strcat(dst_name, dst_wmask);
 819
 820     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name[0]);
 821     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[1], 1, src_name[1]);
 822
 823     if(has_bumpmat) {
 824         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
 825         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", sampler_code);
 826         shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
 827         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", sampler_code);
 828         shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
 829
 830         shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
 831     } else {
 832         shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
 833     }
 834 }
 835
 836 static void pshader_hw_cnd(const struct wined3d_shader_instruction *ins)
 837 {
 838     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 839     SHADER_BUFFER *buffer = ins->buffer;
 840     char dst_wmask[20];
 841     char dst_name[50];
 842     char src_name[3][50];
 843     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
 844     DWORD shift = dst->shift;
 845     BOOL is_color;
 846
 847     /* FIXME: support output modifiers */
 848
 849     /* Handle output register */
 850     shader_arb_get_register_name(ins->shader, dst->register_type,
 851             dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
 852     shader_arb_get_write_mask(ins, dst, dst_wmask);
 853
 854     /* Generate input register names (with modifiers) */
 855     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name[0]);
 856     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[1], 1, src_name[1]);
 857     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[2], 2, src_name[2]);
 858
 859     /* The coissue flag changes the semantic of the cnd instruction in <= 1.3 shaders */
 860     if (ins->reg_maps->shader_version <= WINED3DPS_VERSION(1, 3) && ins->coissue)
 861     {
 862         shader_addline(buffer, "MOV%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1]);
 863     } else {
 864         shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
 865         shader_addline(buffer, "CMP%s %s%s, TMP, %s, %s;\n",
 866                                 sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1], src_name[2]);
 867     }
 868     if (shift != 0)
 869         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 870 }
 871
 872 static void pshader_hw_cmp(const struct wined3d_shader_instruction *ins)
 873 {
 874     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 875     SHADER_BUFFER *buffer = ins->buffer;
 876     char dst_wmask[20];
 877     char dst_name[50];
 878     char src_name[3][50];
 879     DWORD shift = dst->shift;
 880     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
 881     BOOL is_color;
 882
 883     /* FIXME: support output modifiers */
 884
 885     /* Handle output register */
 886     shader_arb_get_register_name(ins->shader, dst->register_type,
 887             dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
 888     shader_arb_get_write_mask(ins, dst, dst_wmask);
 889
 890     /* Generate input register names (with modifiers) */
 891     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name[0]);
 892     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[1], 1, src_name[1]);
 893     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[2], 2, src_name[2]);
 894
 895     shader_addline(buffer, "CMP%s %s%s, %s, %s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
 896                    src_name[0], src_name[2], src_name[1]);
 897
 898     if (shift != 0)
 899         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 900 }
 901
 902 /** Process the WINED3DSIO_DP2ADD instruction in ARB.
 903  * dst = dot2(src0, src1) + src2 */
 904 static void pshader_hw_dp2add(const struct wined3d_shader_instruction *ins)
 905 {
 906     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
 907     SHADER_BUFFER *buffer = ins->buffer;
 908     char dst_wmask[20];
 909     char dst_name[50];
 910     char src_name[3][50];
 911     DWORD shift = dst->shift;
 912     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
 913     BOOL is_color;
 914
 915     shader_arb_get_register_name(ins->shader, dst->register_type,
 916             dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
 917     shader_arb_get_write_mask(ins, dst, dst_wmask);
 918
 919     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name[0]);
 920     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[1], 1, src_name[1]);
 921     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[2], 2, src_name[2]);
 922
 923     /* Emulate a DP2 with a DP3 and 0.0 */
 924     shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
 925     shader_addline(buffer, "MOV TMP.z, 0.0;\n");
 926     shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
 927     shader_addline(buffer, "ADD%s %s%s, TMP2, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[2]);
 928
 929     if (shift != 0)
 930         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
 931 }
 932
 933 /* Map the opcode 1-to-1 to the GL code */
 934 static void shader_hw_map2gl(const struct wined3d_shader_instruction *ins)
 935 {
 936     SHADER_BUFFER *buffer = ins->buffer;
 937     const DWORD *src = ins->src;
 938     const char *instruction;
 939     char arguments[256];
 940     unsigned int i;
 941
 942     switch (ins->handler_idx)
 943     {
 944         case WINED3DSIH_ABS: instruction = "ABS"; break;
 945         case WINED3DSIH_ADD: instruction = "ADD"; break;
 946         case WINED3DSIH_CRS: instruction = "XPD"; break;
 947         case WINED3DSIH_DP3: instruction = "DP3"; break;
 948         case WINED3DSIH_DP4: instruction = "DP4"; break;
 949         case WINED3DSIH_DST: instruction = "DST"; break;
 950         case WINED3DSIH_EXP: instruction = "EX2"; break;
 951         case WINED3DSIH_EXPP: instruction = "EXP"; break;
 952         case WINED3DSIH_FRC: instruction = "FRC"; break;
 953         case WINED3DSIH_LIT: instruction = "LIT"; break;
 954         case WINED3DSIH_LOG: instruction = "LG2"; break;
 955         case WINED3DSIH_LOGP: instruction = "LOG"; break;
 956         case WINED3DSIH_LRP: instruction = "LRP"; break;
 957         case WINED3DSIH_MAD: instruction = "MAD"; break;
 958         case WINED3DSIH_MAX: instruction = "MAX"; break;
 959         case WINED3DSIH_MIN: instruction = "MIN"; break;
 960         case WINED3DSIH_MOV: instruction = "MOV"; break;
 961         case WINED3DSIH_MUL: instruction = "MUL"; break;
 962         case WINED3DSIH_NOP: instruction = "NOP"; break;
 963         case WINED3DSIH_POW: instruction = "POW"; break;
 964         case WINED3DSIH_SGE: instruction = "SGE"; break;
 965         case WINED3DSIH_SLT: instruction = "SLT"; break;
 966         case WINED3DSIH_SUB: instruction = "SUB"; break;
 967         default: instruction = "";
 968             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
 969             break;
 970     }
 971
 972     if (shader_is_pshader_version(ins->reg_maps->shader_version))
 973     {
 974         /* Output token related */
 975         const struct wined3d_shader_dst_param *dst;
 976         char output_rname[256];
 977         char output_wmask[20];
 978         char operands[4][100];
 979         BOOL saturate = FALSE;
 980         BOOL centroid = FALSE;
 981         BOOL partialprecision = FALSE;
 982         const char *modifier;
 983         BOOL is_color;
 984         DWORD shift;
 985
 986         if (!(ins->dst_count + ins->src_count))
 987         {
 988             ERR("Opcode \"%#x\" has no parameters\n", ins->handler_idx);
 989             return;
 990         }
 991         dst = &ins->dst[0];
 992
 993         /* Process modifiers */
 994         if (dst->modifiers)
 995         {
 996             DWORD mask = dst->modifiers;
 997
 998             saturate = mask & WINED3DSPDM_SATURATE;
 999             centroid = mask & WINED3DSPDM_MSAMPCENTROID;
1000             partialprecision = mask & WINED3DSPDM_PARTIALPRECISION;
1001             mask &= ~(WINED3DSPDM_MSAMPCENTROID | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_SATURATE);
1002             if (mask)
1003                 FIXME("Unrecognized modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
1004
1005             if (centroid)
1006                 FIXME("Unhandled modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
1007         }
1008         shift = dst->shift;
1009         modifier = (saturate && !shift) ? "_SAT" : "";
1010
1011         /* Generate input register names (with modifiers) */
1012         for (i = 0; i < ins->src_count; ++i)
1013         {
1014             pshader_gen_input_modifier_line(ins->shader, buffer, src[i], i, operands[i + 1]);
1015         }
1016
1017         /* Handle output register */
1018         shader_arb_get_register_name(ins->shader, dst->register_type,
1019                 dst->register_idx, dst->has_rel_addr, output_rname, &is_color);
1020         strcpy(operands[0], output_rname);
1021         shader_arb_get_write_mask(ins, dst, output_wmask);
1022         strcat(operands[0], output_wmask);
1023
1024         arguments[0] = '\0';
1025         strcat(arguments, operands[0]);
1026         for (i = 0; i < ins->src_count; ++i)
1027         {
1028             strcat(arguments, ", ");
1029             strcat(arguments, operands[i + 1]);
1030         }
1031         shader_addline(buffer, "%s%s %s;\n", instruction, modifier, arguments);
1032
1033         /* A shift requires another line. */
1034         if (shift) pshader_gen_output_modifier_line(buffer, saturate, output_wmask, shift, output_rname);
1035     } else {
1036         /* Note that shader_arb_add_*_param() adds spaces. */
1037
1038         arguments[0] = '\0';
1039         if (ins->dst_count)
1040         {
1041             shader_arb_add_dst_param(ins, &ins->dst[0], arguments);
1042             for (i = 0; i < ins->src_count; ++i)
1043             {
1044                 strcat(arguments, ",");
1045                 shader_arb_add_src_param(ins, src[i], arguments);
1046             }
1047         }
1048         shader_addline(buffer, "%s%s;\n", instruction, arguments);
1049     }
1050 }
1051
1052 static void shader_hw_mov(const struct wined3d_shader_instruction *ins)
1053 {
1054     IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)ins->shader;
1055
1056     if ((WINED3DSHADER_VERSION_MAJOR(ins->reg_maps->shader_version) == 1
1057             && !shader_is_pshader_version(ins->reg_maps->shader_version)
1058             && ins->dst[0].register_type == WINED3DSPR_ADDR)
1059             || ins->handler_idx == WINED3DSIH_MOVA)
1060     {
1061         SHADER_BUFFER *buffer = ins->buffer;
1062         char src0_param[256];
1063
1064         if (ins->handler_idx == WINED3DSIH_MOVA)
1065             FIXME("mova should round\n");
1066
1067         src0_param[0] = '\0';
1068         if (((IWineD3DVertexShaderImpl *)shader)->rel_offset)
1069         {
1070             shader_arb_add_src_param(ins, ins->src[0], src0_param);
1071             shader_addline(buffer, "ADD TMP.x, %s, helper_const.z;\n", src0_param);
1072             shader_addline(buffer, "ARL A0.x, TMP.x;\n");
1073         }
1074         else
1075         {
1076             /* Apple's ARB_vertex_program implementation does not accept an ARL source argument
1077              * with more than one component. Thus replicate the first source argument over all
1078              * 4 components. For example, .xyzw -> .x (or better: .xxxx), .zwxy -> .z, etc)
1079              */
1080             DWORD parm = ins->src[0] & ~(WINED3DVS_SWIZZLE_MASK);
1081             if((ins->src[0] & WINED3DVS_X_W) == WINED3DVS_X_W)
1082                 parm |= WINED3DVS_X_W | WINED3DVS_Y_W | WINED3DVS_Z_W | WINED3DVS_W_W;
1083             else if((ins->src[0] & WINED3DVS_X_Z) == WINED3DVS_X_Z)
1084                 parm |= WINED3DVS_X_Z | WINED3DVS_Y_Z | WINED3DVS_Z_Z | WINED3DVS_W_Z;
1085             else if((ins->src[0] & WINED3DVS_X_Y) == WINED3DVS_X_Y)
1086                 parm |= WINED3DVS_X_Y | WINED3DVS_Y_Y | WINED3DVS_Z_Y | WINED3DVS_W_Y;
1087             else if((ins->src[0] & WINED3DVS_X_X) == WINED3DVS_X_X)
1088                 parm |= WINED3DVS_X_X | WINED3DVS_Y_X | WINED3DVS_Z_X | WINED3DVS_W_X;
1089             shader_arb_add_src_param(ins, parm, src0_param);
1090             shader_addline(buffer, "ARL A0.x, %s;\n", src0_param);
1091         }
1092     }
1093     else
1094     {
1095         shader_hw_map2gl(ins);
1096     }
1097 }
1098
1099 static void pshader_hw_texkill(const struct wined3d_shader_instruction *ins)
1100 {
1101     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1102     DWORD shader_version = ins->reg_maps->shader_version;
1103     SHADER_BUFFER *buffer = ins->buffer;
1104     char reg_dest[40];
1105     BOOL is_color;
1106
1107     /* No swizzles are allowed in d3d's texkill. PS 1.x ignores the 4th component as documented,
1108      * but >= 2.0 honors it(undocumented, but tested by the d3d9 testsuit)
1109      */
1110     shader_arb_get_register_name(ins->shader, dst->register_type,
1111             dst->register_idx, dst->has_rel_addr, reg_dest, &is_color);
1112
1113     if (shader_version >= WINED3DPS_VERSION(2,0))
1114     {
1115         /* The arb backend doesn't claim ps 2.0 support, but try to eat what the app feeds to us */
1116         shader_addline(buffer, "KIL %s;\n", reg_dest);
1117     } else {
1118         /* ARB fp doesn't like swizzles on the parameter of the KIL instruction. To mask the 4th component,
1119          * copy the register into our general purpose TMP variable, overwrite .w and pass TMP to KIL
1120          */
1121         shader_addline(buffer, "MOV TMP, %s;\n", reg_dest);
1122         shader_addline(buffer, "MOV TMP.w, one.w;\n");
1123         shader_addline(buffer, "KIL TMP;\n");
1124     }
1125 }
1126
1127 static void pshader_hw_tex(const struct wined3d_shader_instruction *ins)
1128 {
1129     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1130     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1131     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1132     BOOL is_color;
1133
1134     const DWORD *src = ins->src;
1135     SHADER_BUFFER* buffer = ins->buffer;
1136     DWORD shader_version = ins->reg_maps->shader_version;
1137     BOOL projected = FALSE, bias = FALSE;
1138
1139     char reg_dest[40];
1140     char reg_coord[40];
1141     DWORD reg_sampler_code;
1142
1143     /* All versions have a destination register */
1144     shader_arb_get_register_name(ins->shader, dst->register_type,
1145             dst->register_idx, dst->has_rel_addr, reg_dest, &is_color);
1146
1147     /* 1.0-1.3: Use destination register as coordinate source.
1148        1.4+: Use provided coordinate source register. */
1149    if (shader_version < WINED3DPS_VERSION(1,4))
1150       strcpy(reg_coord, reg_dest);
1151    else
1152       pshader_gen_input_modifier_line(ins->shader, buffer, src[0], 0, reg_coord);
1153
1154   /* 1.0-1.4: Use destination register number as texture code.
1155      2.0+: Use provided sampler number as texure code. */
1156   if (shader_version < WINED3DPS_VERSION(2,0))
1157      reg_sampler_code = dst->register_idx;
1158   else
1159      reg_sampler_code = src[1] & WINED3DSP_REGNUM_MASK;
1160
1161   /* projection flag:
1162    * 1.1, 1.2, 1.3: Use WINED3DTSS_TEXTURETRANSFORMFLAGS
1163    * 1.4: Use WINED3DSPSM_DZ or WINED3DSPSM_DW on src[0]
1164    * 2.0+: Use WINED3DSI_TEXLD_PROJECT on the opcode
1165    */
1166   if (shader_version < WINED3DPS_VERSION(1,4))
1167   {
1168       DWORD flags = 0;
1169       if(reg_sampler_code < MAX_TEXTURES) {
1170         flags = deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1171       }
1172       if (flags & WINED3DTTFF_PROJECTED) {
1173           projected = TRUE;
1174       }
1175   }
1176   else if (shader_version < WINED3DPS_VERSION(2,0))
1177   {
1178       DWORD src_mod = ins->src[0] & WINED3DSP_SRCMOD_MASK;
1179       if (src_mod == WINED3DSPSM_DZ) {
1180           projected = TRUE;
1181       } else if(src_mod == WINED3DSPSM_DW) {
1182           projected = TRUE;
1183       }
1184   } else {
1185       if (ins->flags & WINED3DSI_TEXLD_PROJECT) projected = TRUE;
1186       if (ins->flags & WINED3DSI_TEXLD_BIAS) bias = TRUE;
1187   }
1188   shader_hw_sample(ins, reg_sampler_code, reg_dest, reg_coord, projected, bias);
1189 }
1190
1191 static void pshader_hw_texcoord(const struct wined3d_shader_instruction *ins)
1192 {
1193     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1194     SHADER_BUFFER *buffer = ins->buffer;
1195
1196     char tmp[20];
1197     shader_arb_get_write_mask(ins, dst, tmp);
1198     if (ins->reg_maps->shader_version != WINED3DPS_VERSION(1,4))
1199     {
1200         DWORD reg = dst->register_idx;
1201         shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
1202     } else {
1203         char reg_src[40];
1204
1205         pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, reg_src);
1206         shader_addline(buffer, "MOV R%u%s, %s;\n", dst->register_idx, tmp, reg_src);
1207    }
1208 }
1209
1210 static void pshader_hw_texreg2ar(const struct wined3d_shader_instruction *ins)
1211 {
1212      SHADER_BUFFER *buffer = ins->buffer;
1213      IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1214      IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1215      DWORD flags;
1216
1217      DWORD reg1 = ins->dst[0].register_idx;
1218      char dst_str[8];
1219      char src_str[50];
1220
1221      sprintf(dst_str, "T%u", reg1);
1222      pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_str);
1223      shader_addline(buffer, "MOV TMP.x, %s.w;\n", src_str);
1224      shader_addline(buffer, "MOV TMP.y, %s.x;\n", src_str);
1225      flags = reg1 < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg1][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1226      shader_hw_sample(ins, reg1, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1227 }
1228
1229 static void pshader_hw_texreg2gb(const struct wined3d_shader_instruction *ins)
1230 {
1231      SHADER_BUFFER *buffer = ins->buffer;
1232
1233      DWORD reg1 = ins->dst[0].register_idx;
1234      char dst_str[8];
1235      char src_str[50];
1236
1237      sprintf(dst_str, "T%u", reg1);
1238      pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_str);
1239      shader_addline(buffer, "MOV TMP.x, %s.y;\n", src_str);
1240      shader_addline(buffer, "MOV TMP.y, %s.z;\n", src_str);
1241      shader_hw_sample(ins, reg1, dst_str, "TMP", FALSE, FALSE);
1242 }
1243
1244 static void pshader_hw_texreg2rgb(const struct wined3d_shader_instruction *ins)
1245 {
1246     SHADER_BUFFER *buffer = ins->buffer;
1247     DWORD reg1 = ins->dst[0].register_idx;
1248     char dst_str[8];
1249     char src_str[50];
1250
1251     sprintf(dst_str, "T%u", reg1);
1252     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_str);
1253     shader_hw_sample(ins, reg1, dst_str, src_str, FALSE, FALSE);
1254 }
1255
1256 static void pshader_hw_texbem(const struct wined3d_shader_instruction *ins)
1257 {
1258     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1259     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1260     BOOL has_bumpmat = FALSE;
1261     BOOL has_luminance = FALSE;
1262     BOOL is_color;
1263     int i;
1264
1265     DWORD src = ins->src[0] & WINED3DSP_REGNUM_MASK;
1266     SHADER_BUFFER* buffer = ins->buffer;
1267
1268     char reg_coord[40];
1269     DWORD reg_dest_code;
1270
1271     /* All versions have a destination register */
1272     reg_dest_code = dst->register_idx;
1273     /* Can directly use the name because texbem is only valid for <= 1.3 shaders */
1274     shader_arb_get_register_name(ins->shader, dst->register_type,
1275             dst->register_idx, dst->has_rel_addr, reg_coord, &is_color);
1276
1277     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1278         if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1279                 && reg_dest_code == This->bumpenvmatconst[i].texunit)
1280         {
1281             has_bumpmat = TRUE;
1282             break;
1283         }
1284     }
1285     for(i = 0; i < This->numbumpenvmatconsts; i++) {
1286         if (This->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1287                 && reg_dest_code == This->luminanceconst[i].texunit)
1288         {
1289             has_luminance = TRUE;
1290             break;
1291         }
1292     }
1293
1294     if(has_bumpmat) {
1295         /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
1296
1297         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", reg_dest_code);
1298         shader_addline(buffer, "DP3 TMP.x, TMP2, T%u;\n", src);
1299         shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", reg_dest_code);
1300         shader_addline(buffer, "DP3 TMP.y, TMP2, T%u;\n", src);
1301
1302         /* with projective textures, texbem only divides the static texture coord, not the displacement,
1303          * so we can't let the GL handle this.
1304          */
1305         if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
1306               & WINED3DTTFF_PROJECTED) {
1307             shader_addline(buffer, "RCP TMP2.w, %s.w;\n", reg_coord);
1308             shader_addline(buffer, "MUL TMP2.xy, %s, TMP2.w;\n", reg_coord);
1309             shader_addline(buffer, "ADD TMP.xy, TMP, TMP2;\n");
1310         } else {
1311             shader_addline(buffer, "ADD TMP.xy, TMP, %s;\n", reg_coord);
1312         }
1313
1314         shader_hw_sample(ins, reg_dest_code, reg_coord, "TMP", FALSE, FALSE);
1315
1316         if (ins->handler_idx == WINED3DSIH_TEXBEML && has_luminance)
1317         {
1318             shader_addline(buffer, "MAD TMP, T%u.z, luminance%d.x, luminance%d.y;\n",
1319                            src, reg_dest_code, reg_dest_code);
1320             shader_addline(buffer, "MUL %s, %s, TMP;\n", reg_coord, reg_coord);
1321         }
1322
1323     } else {
1324         DWORD tf;
1325         if(reg_dest_code < MAX_TEXTURES) {
1326             tf = ((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1327         } else {
1328             tf = 0;
1329         }
1330         /* Without a bump matrix loaded, just sample with the unmodified coordinates */
1331         shader_hw_sample(ins, reg_dest_code, reg_coord, reg_coord, tf & WINED3DTTFF_PROJECTED, FALSE);
1332     }
1333 }
1334
1335 static void pshader_hw_texm3x2pad(const struct wined3d_shader_instruction *ins)
1336 {
1337     DWORD reg = ins->dst[0].register_idx;
1338     SHADER_BUFFER *buffer = ins->buffer;
1339     char src0_name[50];
1340
1341     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1342     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
1343 }
1344
1345 static void pshader_hw_texm3x2tex(const struct wined3d_shader_instruction *ins)
1346 {
1347     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1348     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1349     DWORD flags;
1350     DWORD reg = ins->dst[0].register_idx;
1351     SHADER_BUFFER *buffer = ins->buffer;
1352     char dst_str[8];
1353     char src0_name[50];
1354
1355     sprintf(dst_str, "T%u", reg);
1356     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1357     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
1358     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1359     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1360 }
1361
1362 static void pshader_hw_texm3x3pad(const struct wined3d_shader_instruction *ins)
1363 {
1364     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1365     DWORD reg = ins->dst[0].register_idx;
1366     SHADER_BUFFER *buffer = ins->buffer;
1367     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1368     char src0_name[50];
1369
1370     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1371     shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
1372     current_state->texcoord_w[current_state->current_row++] = reg;
1373 }
1374
1375 static void pshader_hw_texm3x3tex(const struct wined3d_shader_instruction *ins)
1376 {
1377     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1378     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1379     DWORD flags;
1380     DWORD reg = ins->dst[0].register_idx;
1381     SHADER_BUFFER *buffer = ins->buffer;
1382     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1383     char dst_str[8];
1384     char src0_name[50];
1385
1386     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1387     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1388
1389     /* Sample the texture using the calculated coordinates */
1390     sprintf(dst_str, "T%u", reg);
1391     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1392     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1393     current_state->current_row = 0;
1394 }
1395
1396 static void pshader_hw_texm3x3vspec(const struct wined3d_shader_instruction *ins)
1397 {
1398     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1399     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1400     DWORD flags;
1401     DWORD reg = ins->dst[0].register_idx;
1402     SHADER_BUFFER *buffer = ins->buffer;
1403     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1404     char dst_str[8];
1405     char src0_name[50];
1406
1407     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1408     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1409
1410     /* Construct the eye-ray vector from w coordinates */
1411     shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
1412     shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
1413     shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
1414
1415     /* Calculate reflection vector
1416      */
1417     shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
1418     /* The .w is ignored when sampling, so I can use TMP2.w to calculate dot(N, N) */
1419     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1420     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1421     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1422     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1423     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
1424
1425     /* Sample the texture using the calculated coordinates */
1426     sprintf(dst_str, "T%u", reg);
1427     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1428     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1429     current_state->current_row = 0;
1430 }
1431
1432 static void pshader_hw_texm3x3spec(const struct wined3d_shader_instruction *ins)
1433 {
1434     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1435     IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1436     DWORD flags;
1437     DWORD reg = ins->dst[0].register_idx;
1438     DWORD reg3 = ins->src[1] & WINED3DSP_REGNUM_MASK;
1439     SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1440     SHADER_BUFFER *buffer = ins->buffer;
1441     char dst_str[8];
1442     char src0_name[50];
1443
1444     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1445     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1446
1447     /* Calculate reflection vector.
1448      *
1449      *               dot(N, E)
1450      * TMP.xyz = 2 * --------- * N - E
1451      *               dot(N, N)
1452      *
1453      * Which normalizes the normal vector
1454      */
1455     shader_addline(buffer, "DP3 TMP.w, TMP, C[%u];\n", reg3);
1456     shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1457     shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1458     shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1459     shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1460     shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -C[%u];\n", reg3);
1461
1462     /* Sample the texture using the calculated coordinates */
1463     sprintf(dst_str, "T%u", reg);
1464     flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1465     shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1466     current_state->current_row = 0;
1467 }
1468
1469 static void pshader_hw_texdepth(const struct wined3d_shader_instruction *ins)
1470 {
1471     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1472     SHADER_BUFFER *buffer = ins->buffer;
1473     char dst_name[50];
1474     BOOL is_color;
1475
1476     /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
1477      * which is essentially an input, is the destination register because it is the first
1478      * parameter. According to the msdn, this must be register r5, but let's keep it more flexible
1479      * here
1480      */
1481     shader_arb_get_register_name(ins->shader, dst->register_type,
1482             dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
1483
1484     /* According to the msdn, the source register(must be r5) is unusable after
1485      * the texdepth instruction, so we're free to modify it
1486      */
1487     shader_addline(buffer, "MIN %s.y, %s.y, one.y;\n", dst_name, dst_name);
1488
1489     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1490      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1491      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1492      */
1493     shader_addline(buffer, "RCP %s.y, %s.y;\n", dst_name, dst_name);
1494     shader_addline(buffer, "MUL TMP.x, %s.x, %s.y;\n", dst_name, dst_name);
1495     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1496     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1497 }
1498
1499 /** Process the WINED3DSIO_TEXDP3TEX instruction in ARB:
1500  * Take a 3-component dot product of the TexCoord[dstreg] and src,
1501  * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1502 static void pshader_hw_texdp3tex(const struct wined3d_shader_instruction *ins)
1503 {
1504     SHADER_BUFFER *buffer = ins->buffer;
1505     DWORD sampler_idx = ins->dst[0].register_idx;
1506     char src0[50];
1507     char dst_str[8];
1508
1509     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0);
1510     shader_addline(buffer, "MOV TMP, 0.0;\n");
1511     shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", sampler_idx, src0);
1512
1513     sprintf(dst_str, "T%u", sampler_idx);
1514     shader_hw_sample(ins, sampler_idx, dst_str, "TMP", FALSE /* Only one coord, can't be projected */, FALSE);
1515 }
1516
1517 /** Process the WINED3DSIO_TEXDP3 instruction in ARB:
1518  * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1519 static void pshader_hw_texdp3(const struct wined3d_shader_instruction *ins)
1520 {
1521     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1522     char src0[50];
1523     char dst_str[50];
1524     char dst_mask[6];
1525     SHADER_BUFFER *buffer = ins->buffer;
1526     BOOL is_color;
1527
1528     /* Handle output register */
1529     shader_arb_get_register_name(ins->shader, dst->register_type,
1530             dst->register_idx, dst->has_rel_addr, dst_str, &is_color);
1531     shader_arb_get_write_mask(ins, dst, dst_mask);
1532
1533     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0);
1534     shader_addline(buffer, "DP3 %s%s, T%u, %s;\n", dst_str, dst_mask, dst->register_idx, src0);
1535
1536     /* TODO: Handle output modifiers */
1537 }
1538
1539 /** Process the WINED3DSIO_TEXM3X3 instruction in ARB
1540  * Perform the 3rd row of a 3x3 matrix multiply */
1541 static void pshader_hw_texm3x3(const struct wined3d_shader_instruction *ins)
1542 {
1543     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1544     SHADER_BUFFER *buffer = ins->buffer;
1545     char dst_str[50];
1546     char dst_mask[6];
1547     char src0[50];
1548     BOOL is_color;
1549
1550     shader_arb_get_register_name(ins->shader, dst->register_type,
1551             dst->register_idx, dst->has_rel_addr, dst_str, &is_color);
1552     shader_arb_get_write_mask(ins, dst, dst_mask);
1553
1554     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0);
1555     shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", dst->register_idx, src0);
1556     shader_addline(buffer, "MOV %s%s, TMP;\n", dst_str, dst_mask);
1557
1558     /* TODO: Handle output modifiers */
1559 }
1560
1561 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in ARB:
1562  * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1563  * Calculate tmp0.y = TexCoord[dstreg] . src.xyz;  (tmp0.x has already been calculated)
1564  * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1565  */
1566 static void pshader_hw_texm3x2depth(const struct wined3d_shader_instruction *ins)
1567 {
1568     SHADER_BUFFER *buffer = ins->buffer;
1569     DWORD dst_reg = ins->dst[0].register_idx;
1570     char src0[50];
1571
1572     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0);
1573     shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", dst_reg, src0);
1574
1575     /* How to deal with the special case dst_name.g == 0? if r != 0, then
1576      * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1577      * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1578      */
1579     shader_addline(buffer, "RCP TMP.y, TMP.y;\n");
1580     shader_addline(buffer, "MUL TMP.x, TMP.x, TMP.y;\n");
1581     shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1582     shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1583 }
1584
1585 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1586     Vertex/Pixel shaders to ARB_vertex_program codes */
1587 static void shader_hw_mnxn(const struct wined3d_shader_instruction *ins)
1588 {
1589     int i;
1590     int nComponents = 0;
1591     struct wined3d_shader_dst_param tmp_dst = {0};
1592     struct wined3d_shader_instruction tmp_ins;
1593
1594     memset(&tmp_ins, 0, sizeof(tmp_ins));
1595
1596     /* Set constants for the temporary argument */
1597     tmp_ins.shader      = ins->shader;
1598     tmp_ins.buffer      = ins->buffer;
1599     tmp_ins.src[0]      = ins->src[0];
1600     tmp_ins.src_addr[0] = ins->src_addr[0];
1601     tmp_ins.src_addr[1] = ins->src_addr[1];
1602     tmp_ins.reg_maps = ins->reg_maps;
1603     tmp_ins.dst_count = 1;
1604     tmp_ins.dst = &tmp_dst;
1605     tmp_ins.src_count = 2;
1606
1607     switch(ins->handler_idx)
1608     {
1609         case WINED3DSIH_M4x4:
1610             nComponents = 4;
1611             tmp_ins.handler_idx = WINED3DSIH_DP4;
1612             break;
1613         case WINED3DSIH_M4x3:
1614             nComponents = 3;
1615             tmp_ins.handler_idx = WINED3DSIH_DP4;
1616             break;
1617         case WINED3DSIH_M3x4:
1618             nComponents = 4;
1619             tmp_ins.handler_idx = WINED3DSIH_DP3;
1620             break;
1621         case WINED3DSIH_M3x3:
1622             nComponents = 3;
1623             tmp_ins.handler_idx = WINED3DSIH_DP3;
1624             break;
1625         case WINED3DSIH_M3x2:
1626             nComponents = 2;
1627             tmp_ins.handler_idx = WINED3DSIH_DP3;
1628             break;
1629         default:
1630             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1631             break;
1632     }
1633
1634     tmp_dst = ins->dst[0];
1635     for (i = 0; i < nComponents; i++) {
1636         tmp_dst.write_mask = WINED3DSP_WRITEMASK_0 << i;
1637         tmp_ins.src[1] = ins->src[1]+i;
1638         shader_hw_map2gl(&tmp_ins);
1639     }
1640 }
1641
1642 static void vshader_hw_rsq_rcp(const struct wined3d_shader_instruction *ins)
1643 {
1644     SHADER_BUFFER *buffer = ins->buffer;
1645     DWORD src = ins->src[0];
1646     DWORD swizzle = (src & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
1647     const char *instruction;
1648
1649     char tmpLine[256];
1650
1651     switch(ins->handler_idx)
1652     {
1653         case WINED3DSIH_RSQ: instruction = "RSQ"; break;
1654         case WINED3DSIH_RCP: instruction = "RCP"; break;
1655         default: instruction = "";
1656             FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1657             break;
1658     }
1659
1660     strcpy(tmpLine, instruction);
1661     shader_arb_add_dst_param(ins, &ins->dst[0], tmpLine); /* Destination */
1662     strcat(tmpLine, ",");
1663     shader_arb_add_src_param(ins, src, tmpLine);
1664     if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) == swizzle) {
1665         /* Dx sdk says .x is used if no swizzle is given, but our test shows that
1666          * .w is used
1667          */
1668         strcat(tmpLine, ".w");
1669     }
1670
1671     shader_addline(buffer, "%s;\n", tmpLine);
1672 }
1673
1674 static void shader_hw_nrm(const struct wined3d_shader_instruction *ins)
1675 {
1676     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1677     SHADER_BUFFER *buffer = ins->buffer;
1678     char dst_name[50];
1679     char src_name[50];
1680     char dst_wmask[20];
1681     DWORD shift = dst->shift;
1682     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1683     BOOL is_color;
1684
1685     shader_arb_get_register_name(ins->shader, dst->register_type,
1686             dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
1687     shader_arb_get_write_mask(ins, dst, dst_wmask);
1688
1689     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name);
1690     shader_addline(buffer, "DP3 TMP, %s, %s;\n", src_name, src_name);
1691     shader_addline(buffer, "RSQ TMP, TMP.x;\n");
1692     /* dst.w = src[0].w * 1 / (src.x^2 + src.y^2 + src.z^2)^(1/2) according to msdn*/
1693     shader_addline(buffer, "MUL%s %s%s, %s, TMP;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1694                    src_name);
1695
1696     if (shift != 0)
1697         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1698 }
1699
1700 static void shader_hw_sincos(const struct wined3d_shader_instruction *ins)
1701 {
1702     /* This instruction exists in ARB, but the d3d instruction takes two extra parameters which
1703      * must contain fixed constants. So we need a separate function to filter those constants and
1704      * can't use map2gl
1705      */
1706     const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1707     SHADER_BUFFER *buffer = ins->buffer;
1708     char dst_name[50];
1709     char src_name[50];
1710     char dst_wmask[20];
1711     DWORD shift = dst->shift;
1712     BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1713     BOOL is_color;
1714
1715     shader_arb_get_register_name(ins->shader, dst->register_type,
1716             dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
1717     shader_arb_get_write_mask(ins, dst, dst_wmask);
1718
1719     pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name);
1720     shader_addline(buffer, "SCS%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1721                    src_name);
1722
1723     if (shift != 0)
1724         pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1725
1726 }
1727
1728 static GLuint create_arb_blt_vertex_program(const WineD3D_GL_Info *gl_info)
1729 {
1730     GLuint program_id = 0;
1731     const char *blt_vprogram =
1732         "!!ARBvp1.0\n"
1733         "PARAM c[1] = { { 1, 0.5 } };\n"
1734         "MOV result.position, vertex.position;\n"
1735         "MOV result.color, c[0].x;\n"
1736         "MOV result.texcoord[0], vertex.texcoord[0];\n"
1737         "END\n";
1738
1739     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1740     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
1741     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1742
1743     if (glGetError() == GL_INVALID_OPERATION) {
1744         GLint pos;
1745         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1746         FIXME("Vertex program error at position %d: %s\n", pos,
1747             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1748     }
1749
1750     return program_id;
1751 }
1752
1753 static GLuint create_arb_blt_fragment_program(const WineD3D_GL_Info *gl_info, enum tex_types tex_type)
1754 {
1755     GLuint program_id = 0;
1756     static const char * const blt_fprograms[tex_type_count] =
1757     {
1758         /* tex_1d */
1759         NULL,
1760         /* tex_2d */
1761         "!!ARBfp1.0\n"
1762         "TEMP R0;\n"
1763         "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1764         "MOV result.depth.z, R0.x;\n"
1765         "END\n",
1766         /* tex_3d */
1767         NULL,
1768         /* tex_cube */
1769         "!!ARBfp1.0\n"
1770         "TEMP R0;\n"
1771         "TEX R0.x, fragment.texcoord[0], texture[0], CUBE;\n"
1772         "MOV result.depth.z, R0.x;\n"
1773         "END\n",
1774         /* tex_rect */
1775         "!!ARBfp1.0\n"
1776         "TEMP R0;\n"
1777         "TEX R0.x, fragment.texcoord[0], texture[0], RECT;\n"
1778         "MOV result.depth.z, R0.x;\n"
1779         "END\n",
1780     };
1781
1782     if (!blt_fprograms[tex_type])
1783     {
1784         FIXME("tex_type %#x not supported\n", tex_type);
1785         tex_type = tex_2d;
1786     }
1787
1788     GL_EXTCALL(glGenProgramsARB(1, &program_id));
1789     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1790     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprograms[tex_type]), blt_fprograms[tex_type]));
1791
1792     if (glGetError() == GL_INVALID_OPERATION) {
1793         GLint pos;
1794         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1795         FIXME("Fragment program error at position %d: %s\n", pos,
1796             debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1797     }
1798
1799     return program_id;
1800 }
1801
1802 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1803     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1804     struct shader_arb_priv *priv = This->shader_priv;
1805     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1806
1807     if (useVS) {
1808         struct vs_compile_args compile_args;
1809
1810         TRACE("Using vertex shader\n");
1811         find_vs_compile_args((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, This->stateBlock, &compile_args);
1812         priv->current_vprogram_id = find_gl_vshader((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, &compile_args);
1813
1814         /* Bind the vertex program */
1815         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1816         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id);");
1817
1818         /* Enable OpenGL vertex programs */
1819         glEnable(GL_VERTEX_PROGRAM_ARB);
1820         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1821         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1822     } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1823         priv->current_vprogram_id = 0;
1824         glDisable(GL_VERTEX_PROGRAM_ARB);
1825         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1826     }
1827
1828     if (usePS) {
1829         struct ps_compile_args compile_args;
1830         TRACE("Using pixel shader\n");
1831         find_ps_compile_args((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader, This->stateBlock, &compile_args);
1832         priv->current_fprogram_id = find_gl_pshader((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader,
1833                                                     &compile_args);
1834
1835         /* Bind the fragment program */
1836         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1837         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id);");
1838
1839         if(!priv->use_arbfp_fixed_func) {
1840             /* Enable OpenGL fragment programs */
1841             glEnable(GL_FRAGMENT_PROGRAM_ARB);
1842             checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1843         }
1844         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1845     } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM) && !priv->use_arbfp_fixed_func) {
1846         /* Disable only if we're not using arbfp fixed function fragment processing. If this is used,
1847          * keep GL_FRAGMENT_PROGRAM_ARB enabled, and the fixed function pipeline will bind the fixed function
1848          * replacement shader
1849          */
1850         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1851         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1852         priv->current_fprogram_id = 0;
1853     }
1854 }
1855
1856 static void shader_arb_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {
1857     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1858     struct shader_arb_priv *priv = This->shader_priv;
1859     GLuint *blt_fprogram = &priv->depth_blt_fprogram_id[tex_type];
1860     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1861
1862     if (!priv->depth_blt_vprogram_id) priv->depth_blt_vprogram_id = create_arb_blt_vertex_program(gl_info);
1863     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->depth_blt_vprogram_id));
1864     glEnable(GL_VERTEX_PROGRAM_ARB);
1865
1866     if (!*blt_fprogram) *blt_fprogram = create_arb_blt_fragment_program(gl_info, tex_type);
1867     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, *blt_fprogram));
1868     glEnable(GL_FRAGMENT_PROGRAM_ARB);
1869 }
1870
1871 static void shader_arb_deselect_depth_blt(IWineD3DDevice *iface) {
1872     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1873     struct shader_arb_priv *priv = This->shader_priv;
1874     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1875
1876     if (priv->current_vprogram_id) {
1877         GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1878         checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1879
1880         glEnable(GL_VERTEX_PROGRAM_ARB);
1881         checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1882
1883         TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1884     } else {
1885         glDisable(GL_VERTEX_PROGRAM_ARB);
1886         checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1887     }
1888
1889     if (priv->current_fprogram_id) {
1890         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1891         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1892
1893         glEnable(GL_FRAGMENT_PROGRAM_ARB);
1894         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1895
1896         TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1897     } else {
1898         glDisable(GL_FRAGMENT_PROGRAM_ARB);
1899         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1900     }
1901 }
1902
1903 static void shader_arb_destroy(IWineD3DBaseShader *iface) {
1904     IWineD3DBaseShaderImpl *baseShader = (IWineD3DBaseShaderImpl *) iface;
1905     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)baseShader->baseShader.device)->adapter->gl_info;
1906
1907     if (shader_is_pshader_version(baseShader->baseShader.reg_maps.shader_version))
1908     {
1909         IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *) iface;
1910         UINT i;
1911
1912         ENTER_GL();
1913         for(i = 0; i < This->num_gl_shaders; i++) {
1914             GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1915             checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1916         }
1917         LEAVE_GL();
1918         HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1919         This->gl_shaders = NULL;
1920         This->num_gl_shaders = 0;
1921         This->shader_array_size = 0;
1922     } else {
1923         IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *) iface;
1924         UINT i;
1925
1926         ENTER_GL();
1927         for(i = 0; i < This->num_gl_shaders; i++) {
1928             GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1929             checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1930         }
1931         LEAVE_GL();
1932         HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1933         This->gl_shaders = NULL;
1934         This->num_gl_shaders = 0;
1935         This->shader_array_size = 0;
1936     }
1937 }
1938
1939 static HRESULT shader_arb_alloc(IWineD3DDevice *iface) {
1940     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1941     This->shader_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
1942     return WINED3D_OK;
1943 }
1944
1945 static void shader_arb_free(IWineD3DDevice *iface) {
1946     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1947     const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1948     struct shader_arb_priv *priv = This->shader_priv;
1949     int i;
1950
1951     if(priv->depth_blt_vprogram_id) {
1952         GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_vprogram_id));
1953     }
1954     for (i = 0; i < tex_type_count; ++i) {
1955         if (priv->depth_blt_fprogram_id[i]) {
1956             GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_fprogram_id[i]));
1957         }
1958     }
1959
1960     HeapFree(GetProcessHeap(), 0, This->shader_priv);
1961 }
1962
1963 static BOOL shader_arb_dirty_const(IWineD3DDevice *iface) {
1964     return TRUE;
1965 }
1966
1967 static void arbfp_add_sRGB_correction(SHADER_BUFFER *buffer, const char *fragcolor, const char *tmp1,
1968                                       const char *tmp2, const char *tmp3, const char *tmp4) {
1969     /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
1970
1971     /* Calculate the > 0.0031308 case */
1972     shader_addline(buffer, "POW %s.x, %s.x, srgb_pow.x;\n", tmp1, fragcolor);
1973     shader_addline(buffer, "POW %s.y, %s.y, srgb_pow.y;\n", tmp1, fragcolor);
1974     shader_addline(buffer, "POW %s.z, %s.z, srgb_pow.z;\n", tmp1, fragcolor);
1975     shader_addline(buffer, "MUL %s, %s, srgb_mul_hi;\n", tmp1, tmp1);
1976     shader_addline(buffer, "SUB %s, %s, srgb_sub_hi;\n", tmp1, tmp1);
1977     /* Calculate the < case */
1978     shader_addline(buffer, "MUL %s, srgb_mul_low, %s;\n", tmp2, fragcolor);
1979     /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
1980     shader_addline(buffer, "SLT %s, srgb_comparison, %s;\n", tmp3, fragcolor);
1981     shader_addline(buffer, "SGE %s, srgb_comparison, %s;\n", tmp4, fragcolor);
1982     /* Store the components > 0.0031308 in the destination */
1983     shader_addline(buffer, "MUL %s, %s, %s;\n", fragcolor, tmp1, tmp3);
1984     /* Add the components that are < 0.0031308 */
1985     shader_addline(buffer, "MAD result.color.xyz, %s, %s, %s;\n", tmp2, tmp4, fragcolor);
1986     /* [0.0;1.0] clamping. Not needed, this is done implicitly */
1987 }
1988
1989 static GLuint shader_arb_generate_pshader(IWineD3DPixelShader *iface, SHADER_BUFFER *buffer, const struct ps_compile_args *args) {
1990     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1991     const shader_reg_maps* reg_maps = &This->baseShader.reg_maps;
1992     CONST DWORD *function = This->baseShader.function;
1993     DWORD shader_version = reg_maps->shader_version;
1994     const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
1995     const local_constant *lconst;
1996     GLuint retval;
1997     const char *fragcolor;
1998
1999     /*  Create the hw ARB shader */
2000     shader_addline(buffer, "!!ARBfp1.0\n");
2001
2002     if (shader_version < WINED3DPS_VERSION(3,0)) {
2003         switch(args->fog) {
2004             case FOG_OFF:
2005                 break;
2006             case FOG_LINEAR:
2007                 shader_addline(buffer, "OPTION ARB_fog_linear;\n");
2008                 break;
2009             case FOG_EXP:
2010                 shader_addline(buffer, "OPTION ARB_fog_exp;\n");
2011                 break;
2012             case FOG_EXP2:
2013                 shader_addline(buffer, "OPTION ARB_fog_exp2;\n");
2014                 break;
2015         }
2016     }
2017
2018     shader_addline(buffer, "TEMP TMP;\n");     /* Used in matrix ops */
2019     shader_addline(buffer, "TEMP TMP2;\n");    /* Used in matrix ops */
2020     shader_addline(buffer, "TEMP TA;\n");      /* Used for modifiers */
2021     shader_addline(buffer, "TEMP TB;\n");      /* Used for modifiers */
2022     shader_addline(buffer, "TEMP TC;\n");      /* Used for modifiers */
2023     shader_addline(buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
2024     shader_addline(buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
2025     shader_addline(buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
2026
2027     if (shader_version < WINED3DPS_VERSION(2,0)) {
2028         fragcolor = "R0";
2029     } else {
2030         shader_addline(buffer, "TEMP TMP_COLOR;\n");
2031         fragcolor = "TMP_COLOR";
2032     }
2033
2034     /* Base Declarations */
2035     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
2036
2037     /* Base Shader Body */
2038     shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
2039
2040     if(args->srgb_correction) {
2041         arbfp_add_sRGB_correction(buffer, fragcolor, "TMP", "TMP2", "TA", "TB");
2042     }
2043     shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
2044     shader_addline(buffer, "END\n");
2045
2046     /* TODO: change to resource.glObjectHandle or something like that */
2047     GL_EXTCALL(glGenProgramsARB(1, &retval));
2048
2049     TRACE("Creating a hw pixel shader, prg=%d\n", retval);
2050     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, retval));
2051
2052     TRACE("Created hw pixel shader, prg=%d\n", retval);
2053     /* Create the program and check for errors */
2054     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2055                buffer->bsize, buffer->buffer));
2056
2057     if (glGetError() == GL_INVALID_OPERATION) {
2058         GLint errPos;
2059         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2060         FIXME("HW PixelShader Error at position %d: %s\n",
2061               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2062         retval = 0;
2063     }
2064
2065     /* Load immediate constants */
2066     if(!This->baseShader.load_local_constsF) {
2067         LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2068             const float *value = (const float *)lconst->value;
2069             GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, lconst->idx, value));
2070             checkGLcall("glProgramLocalParameter4fvARB");
2071         }
2072     }
2073
2074     return retval;
2075 }
2076
2077 static GLuint shader_arb_generate_vshader(IWineD3DVertexShader *iface, SHADER_BUFFER *buffer, const struct vs_compile_args *args) {
2078     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
2079     const shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
2080     CONST DWORD *function = This->baseShader.function;
2081     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)This->baseShader.device;
2082     const WineD3D_GL_Info *gl_info = &device->adapter->gl_info;
2083     const local_constant *lconst;
2084     GLuint ret;
2085
2086     /*  Create the hw ARB shader */
2087     shader_addline(buffer, "!!ARBvp1.0\n");
2088     shader_addline(buffer, "PARAM helper_const = { 2.0, -1.0, %d.0, 0.0 };\n", This->rel_offset);
2089
2090     /* Mesa supports only 95 constants */
2091     if (GL_VEND(MESA) || GL_VEND(WINE))
2092         This->baseShader.limits.constant_float =
2093                 min(95, This->baseShader.limits.constant_float);
2094
2095     shader_addline(buffer, "TEMP TMP;\n");
2096
2097     /* Base Declarations */
2098     shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
2099
2100     /* We need a constant to fixup the final position */
2101     shader_addline(buffer, "PARAM posFixup = program.env[%d];\n", ARB_SHADER_PRIVCONST_POS);
2102
2103     /* Initialize output parameters. GL_ARB_vertex_program does not require special initialization values
2104      * for output parameters. D3D in theory does not do that either, but some applications depend on a
2105      * proper initialization of the secondary color, and programs using the fixed function pipeline without
2106      * a replacement shader depend on the texcoord.w being set properly.
2107      *
2108      * GL_NV_vertex_program defines that all output values are initialized to {0.0, 0.0, 0.0, 1.0}. This
2109      * assertion is in effect even when using GL_ARB_vertex_program without any NV specific additions. So
2110      * skip this if NV_vertex_program is supported. Otherwise, initialize the secondary color. For the tex-
2111      * coords, we have a flag in the opengl caps. Many cards do not require the texcoord being set, and
2112      * this can eat a number of instructions, so skip it unless this cap is set as well
2113      */
2114     if(!GL_SUPPORT(NV_VERTEX_PROGRAM)) {
2115         shader_addline(buffer, "MOV result.color.secondary, -helper_const.wwwy;\n");
2116
2117         if((GLINFO_LOCATION).set_texcoord_w && !device->frag_pipe->ffp_proj_control) {
2118             int i;
2119             for(i = 0; i < min(8, MAX_REG_TEXCRD); i++) {
2120                 if(This->baseShader.reg_maps.texcoord_mask[i] != 0 &&
2121                 This->baseShader.reg_maps.texcoord_mask[i] != WINED3DSP_WRITEMASK_ALL) {
2122                     shader_addline(buffer, "MOV result.texcoord[%u].w, -helper_const.y;\n", i);
2123                 }
2124             }
2125         }
2126     }
2127
2128     /* Base Shader Body */
2129     shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
2130
2131     /* The D3DRS_FOGTABLEMODE render state defines if the shader-generated fog coord is used
2132      * or if the fragment depth is used. If the fragment depth is used(FOGTABLEMODE != NONE),
2133      * the fog frag coord is thrown away. If the fog frag coord is used, but not written by
2134      * the shader, it is set to 0.0(fully fogged, since start = 1.0, end = 0.0)
2135      */
2136     if(args->fog_src == VS_FOG_Z) {
2137         shader_addline(buffer, "MOV result.fogcoord, TMP_OUT.z;\n");
2138     } else if (!reg_maps->fog) {
2139         shader_addline(buffer, "MOV result.fogcoord, 0.0;\n");
2140     }
2141
2142     /* Write the final position.
2143      *
2144      * OpenGL coordinates specify the center of the pixel while d3d coords specify
2145      * the corner. The offsets are stored in z and w in posFixup. posFixup.y contains
2146      * 1.0 or -1.0 to turn the rendering upside down for offscreen rendering. PosFixup.x
2147      * contains 1.0 to allow a mad, but arb vs swizzles are too restricted for that.
2148      */
2149     shader_addline(buffer, "MUL TMP, posFixup, TMP_OUT.w;\n");
2150     shader_addline(buffer, "ADD TMP_OUT.x, TMP_OUT.x, TMP.z;\n");
2151     shader_addline(buffer, "MAD TMP_OUT.y, TMP_OUT.y, posFixup.y, TMP.w;\n");
2152
2153     /* Z coord [0;1]->[-1;1] mapping, see comment in transform_projection in state.c
2154      * and the glsl equivalent
2155      */
2156     shader_addline(buffer, "MAD TMP_OUT.z, TMP_OUT.z, helper_const.x, -TMP_OUT.w;\n");
2157
2158     shader_addline(buffer, "MOV result.position, TMP_OUT;\n");
2159
2160     shader_addline(buffer, "END\n");
2161
2162     /* TODO: change to resource.glObjectHandle or something like that */
2163     GL_EXTCALL(glGenProgramsARB(1, &ret));
2164
2165     TRACE("Creating a hw vertex shader, prg=%d\n", ret);
2166     GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, ret));
2167
2168     TRACE("Created hw vertex shader, prg=%d\n", ret);
2169     /* Create the program and check for errors */
2170     GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2171                buffer->bsize, buffer->buffer));
2172
2173     if (glGetError() == GL_INVALID_OPERATION) {
2174         GLint errPos;
2175         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2176         FIXME("HW VertexShader Error at position %d: %s\n",
2177               errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2178         ret = -1;
2179     } else {
2180         /* Load immediate constants */
2181         if(!This->baseShader.load_local_constsF) {
2182             LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2183                 const float *value = (const float *)lconst->value;
2184                 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, lconst->idx, value));
2185             }
2186         }
2187     }
2188     return ret;
2189 }
2190
2191 static void shader_arb_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
2192 {
2193     /* We don't have an ARB fixed function pipeline yet, so let the none backend set its caps,
2194      * then overwrite the shader specific ones
2195      */
2196     none_shader_backend.shader_get_caps(devtype, gl_info, pCaps);
2197
2198     if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
2199         pCaps->VertexShaderVersion = WINED3DVS_VERSION(1,1);
2200         TRACE_(d3d_caps)("Hardware vertex shader version 1.1 enabled (ARB_PROGRAM)\n");
2201         pCaps->MaxVertexShaderConst = GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS;
2202     }
2203
2204     if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
2205         pCaps->PixelShaderVersion    = WINED3DPS_VERSION(1,4);
2206         pCaps->PixelShader1xMaxValue = 8.0;
2207         TRACE_(d3d_caps)("Hardware pixel shader version 1.4 enabled (ARB_PROGRAM)\n");
2208         pCaps->MaxPixelShaderConst = GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS;
2209     }
2210 }
2211
2212 static BOOL shader_arb_color_fixup_supported(struct color_fixup_desc fixup)
2213 {
2214     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
2215     {
2216         TRACE("Checking support for color_fixup:\n");
2217         dump_color_fixup_desc(fixup);
2218     }
2219
2220     /* We support everything except YUV conversions. */
2221     if (!is_yuv_fixup(fixup))
2222     {
2223         TRACE("[OK]\n");
2224         return TRUE;
2225     }
2226
2227     TRACE("[FAILED]\n");
2228     return FALSE;
2229 }
2230
2231 static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABLE_SIZE] =
2232 {
2233     /* WINED3DSIH_ABS           */ shader_hw_map2gl,
2234     /* WINED3DSIH_ADD           */ shader_hw_map2gl,
2235     /* WINED3DSIH_BEM           */ pshader_hw_bem,
2236     /* WINED3DSIH_BREAK         */ NULL,
2237     /* WINED3DSIH_BREAKC        */ NULL,
2238     /* WINED3DSIH_BREAKP        */ NULL,
2239     /* WINED3DSIH_CALL          */ NULL,
2240     /* WINED3DSIH_CALLNZ        */ NULL,
2241     /* WINED3DSIH_CMP           */ pshader_hw_cmp,
2242     /* WINED3DSIH_CND           */ pshader_hw_cnd,
2243     /* WINED3DSIH_CRS           */ shader_hw_map2gl,
2244     /* WINED3DSIH_DCL           */ NULL,
2245     /* WINED3DSIH_DEF           */ NULL,
2246     /* WINED3DSIH_DEFB          */ NULL,
2247     /* WINED3DSIH_DEFI          */ NULL,
2248     /* WINED3DSIH_DP2ADD        */ pshader_hw_dp2add,
2249     /* WINED3DSIH_DP3           */ shader_hw_map2gl,
2250     /* WINED3DSIH_DP4           */ shader_hw_map2gl,
2251     /* WINED3DSIH_DST           */ shader_hw_map2gl,
2252     /* WINED3DSIH_DSX           */ NULL,
2253     /* WINED3DSIH_DSY           */ NULL,
2254     /* WINED3DSIH_ELSE          */ NULL,
2255     /* WINED3DSIH_ENDIF         */ NULL,
2256     /* WINED3DSIH_ENDLOOP       */ NULL,
2257     /* WINED3DSIH_ENDREP        */ NULL,
2258     /* WINED3DSIH_EXP           */ shader_hw_map2gl,
2259     /* WINED3DSIH_EXPP          */ shader_hw_map2gl,
2260     /* WINED3DSIH_FRC           */ shader_hw_map2gl,
2261     /* WINED3DSIH_IF            */ NULL,
2262     /* WINED3DSIH_IFC           */ NULL,
2263     /* WINED3DSIH_LABEL         */ NULL,
2264     /* WINED3DSIH_LIT           */ shader_hw_map2gl,
2265     /* WINED3DSIH_LOG           */ shader_hw_map2gl,
2266     /* WINED3DSIH_LOGP          */ shader_hw_map2gl,
2267     /* WINED3DSIH_LOOP          */ NULL,
2268     /* WINED3DSIH_LRP           */ shader_hw_map2gl,
2269     /* WINED3DSIH_M3x2          */ shader_hw_mnxn,
2270     /* WINED3DSIH_M3x3          */ shader_hw_mnxn,
2271     /* WINED3DSIH_M3x4          */ shader_hw_mnxn,
2272     /* WINED3DSIH_M4x3          */ shader_hw_mnxn,
2273     /* WINED3DSIH_M4x4          */ shader_hw_mnxn,
2274     /* WINED3DSIH_MAD           */ shader_hw_map2gl,
2275     /* WINED3DSIH_MAX           */ shader_hw_map2gl,
2276     /* WINED3DSIH_MIN           */ shader_hw_map2gl,
2277     /* WINED3DSIH_MOV           */ shader_hw_mov,
2278     /* WINED3DSIH_MOVA          */ shader_hw_mov,
2279     /* WINED3DSIH_MUL           */ shader_hw_map2gl,
2280     /* WINED3DSIH_NOP           */ shader_hw_map2gl,
2281     /* WINED3DSIH_NRM           */ shader_hw_nrm,
2282     /* WINED3DSIH_PHASE         */ NULL,
2283     /* WINED3DSIH_POW           */ shader_hw_map2gl,
2284     /* WINED3DSIH_RCP           */ vshader_hw_rsq_rcp,
2285     /* WINED3DSIH_REP           */ NULL,
2286     /* WINED3DSIH_RET           */ NULL,
2287     /* WINED3DSIH_RSQ           */ vshader_hw_rsq_rcp,
2288     /* WINED3DSIH_SETP          */ NULL,
2289     /* WINED3DSIH_SGE           */ shader_hw_map2gl,
2290     /* WINED3DSIH_SGN           */ NULL,
2291     /* WINED3DSIH_SINCOS        */ shader_hw_sincos,
2292     /* WINED3DSIH_SLT           */ shader_hw_map2gl,
2293     /* WINED3DSIH_SUB           */ shader_hw_map2gl,
2294     /* WINED3DSIH_TEX           */ pshader_hw_tex,
2295     /* WINED3DSIH_TEXBEM        */ pshader_hw_texbem,
2296     /* WINED3DSIH_TEXBEML       */ pshader_hw_texbem,
2297     /* WINED3DSIH_TEXCOORD      */ pshader_hw_texcoord,
2298     /* WINED3DSIH_TEXDEPTH      */ pshader_hw_texdepth,
2299     /* WINED3DSIH_TEXDP3        */ pshader_hw_texdp3,
2300     /* WINED3DSIH_TEXDP3TEX     */ pshader_hw_texdp3tex,
2301     /* WINED3DSIH_TEXKILL       */ pshader_hw_texkill,
2302     /* WINED3DSIH_TEXLDD        */ NULL,
2303     /* WINED3DSIH_TEXLDL        */ NULL,
2304     /* WINED3DSIH_TEXM3x2DEPTH  */ pshader_hw_texm3x2depth,
2305     /* WINED3DSIH_TEXM3x2PAD    */ pshader_hw_texm3x2pad,
2306     /* WINED3DSIH_TEXM3x2TEX    */ pshader_hw_texm3x2tex,
2307     /* WINED3DSIH_TEXM3x3       */ pshader_hw_texm3x3,
2308     /* WINED3DSIH_TEXM3x3DIFF   */ NULL,
2309     /* WINED3DSIH_TEXM3x3PAD    */ pshader_hw_texm3x3pad,
2310     /* WINED3DSIH_TEXM3x3SPEC   */ pshader_hw_texm3x3spec,
2311     /* WINED3DSIH_TEXM3x3TEX    */ pshader_hw_texm3x3tex,
2312     /* WINED3DSIH_TEXM3x3VSPEC  */ pshader_hw_texm3x3vspec,
2313     /* WINED3DSIH_TEXREG2AR     */ pshader_hw_texreg2ar,
2314     /* WINED3DSIH_TEXREG2GB     */ pshader_hw_texreg2gb,
2315     /* WINED3DSIH_TEXREG2RGB    */ pshader_hw_texreg2rgb,
2316 };
2317
2318 const shader_backend_t arb_program_shader_backend = {
2319     shader_arb_instruction_handler_table,
2320     shader_arb_select,
2321     shader_arb_select_depth_blt,
2322     shader_arb_deselect_depth_blt,
2323     shader_arb_update_float_vertex_constants,
2324     shader_arb_update_float_pixel_constants,
2325     shader_arb_load_constants,
2326     shader_arb_load_np2fixup_constants,
2327     shader_arb_destroy,
2328     shader_arb_alloc,
2329     shader_arb_free,
2330     shader_arb_dirty_const,
2331     shader_arb_generate_pshader,
2332     shader_arb_generate_vshader,
2333     shader_arb_get_caps,
2334     shader_arb_color_fixup_supported,
2335 };
2336
2337 /* ARB_fragment_program fixed function pipeline replacement definitions */
2338 #define ARB_FFP_CONST_TFACTOR           0
2339 #define ARB_FFP_CONST_SPECULAR_ENABLE   ((ARB_FFP_CONST_TFACTOR) + 1)
2340 #define ARB_FFP_CONST_CONSTANT(i)       ((ARB_FFP_CONST_SPECULAR_ENABLE) + 1 + i)
2341 #define ARB_FFP_CONST_BUMPMAT(i)        ((ARB_FFP_CONST_CONSTANT(7)) + 1 + i)
2342 #define ARB_FFP_CONST_LUMINANCE(i)      ((ARB_FFP_CONST_BUMPMAT(7)) + 1 + i)
2343
2344 struct arbfp_ffp_desc
2345 {
2346     struct ffp_frag_desc parent;
2347     GLuint shader;
2348     unsigned int num_textures_used;
2349 };
2350
2351 static void arbfp_enable(IWineD3DDevice *iface, BOOL enable) {
2352     if(enable) {
2353         glEnable(GL_FRAGMENT_PROGRAM_ARB);
2354         checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
2355     } else {
2356         glDisable(GL_FRAGMENT_PROGRAM_ARB);
2357         checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
2358     }
2359 }
2360
2361 static HRESULT arbfp_alloc(IWineD3DDevice *iface) {
2362     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2363     struct shader_arb_priv *priv;
2364     /* Share private data between the shader backend and the pipeline replacement, if both
2365      * are the arb implementation. This is needed to figure out whether ARBfp should be disabled
2366      * if no pixel shader is bound or not
2367      */
2368     if(This->shader_backend == &arb_program_shader_backend) {
2369         This->fragment_priv = This->shader_priv;
2370     } else {
2371         This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
2372         if(!This->fragment_priv) return E_OUTOFMEMORY;
2373     }
2374     priv = This->fragment_priv;
2375     priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
2376     priv->use_arbfp_fixed_func = TRUE;
2377     return WINED3D_OK;
2378 }
2379
2380 static void arbfp_free_ffpshader(void *value, void *gli) {
2381     const WineD3D_GL_Info *gl_info = gli;
2382     struct arbfp_ffp_desc *entry_arb = value;
2383
2384     ENTER_GL();
2385     GL_EXTCALL(glDeleteProgramsARB(1, &entry_arb->shader));
2386     checkGLcall("glDeleteProgramsARB(1, &entry_arb->shader)");
2387     HeapFree(GetProcessHeap(), 0, entry_arb);
2388     LEAVE_GL();
2389 }
2390
2391 static void arbfp_free(IWineD3DDevice *iface) {
2392     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2393     struct shader_arb_priv *priv = This->fragment_priv;
2394
2395     hash_table_destroy(priv->fragment_shaders, arbfp_free_ffpshader, &This->adapter->gl_info);
2396     priv->use_arbfp_fixed_func = FALSE;
2397
2398     if(This->shader_backend != &arb_program_shader_backend) {
2399         HeapFree(GetProcessHeap(), 0, This->fragment_priv);
2400     }
2401 }
2402
2403 static void arbfp_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
2404 {
2405     caps->TextureOpCaps =  WINED3DTEXOPCAPS_DISABLE                     |
2406                            WINED3DTEXOPCAPS_SELECTARG1                  |
2407                            WINED3DTEXOPCAPS_SELECTARG2                  |
2408                            WINED3DTEXOPCAPS_MODULATE4X                  |
2409                            WINED3DTEXOPCAPS_MODULATE2X                  |
2410                            WINED3DTEXOPCAPS_MODULATE                    |
2411                            WINED3DTEXOPCAPS_ADDSIGNED2X                 |
2412                            WINED3DTEXOPCAPS_ADDSIGNED                   |
2413                            WINED3DTEXOPCAPS_ADD                         |
2414                            WINED3DTEXOPCAPS_SUBTRACT                    |
2415                            WINED3DTEXOPCAPS_ADDSMOOTH                   |
2416                            WINED3DTEXOPCAPS_BLENDCURRENTALPHA           |
2417                            WINED3DTEXOPCAPS_BLENDFACTORALPHA            |
2418                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHA           |
2419                            WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA           |
2420                            WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM         |
2421                            WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR      |
2422                            WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA      |
2423                            WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA   |
2424                            WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR   |
2425                            WINED3DTEXOPCAPS_DOTPRODUCT3                 |
2426                            WINED3DTEXOPCAPS_MULTIPLYADD                 |
2427                            WINED3DTEXOPCAPS_LERP                        |
2428                            WINED3DTEXOPCAPS_BUMPENVMAP                  |
2429                            WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE;
2430
2431     /* TODO: Implement WINED3DTEXOPCAPS_PREMODULATE */
2432
2433     caps->MaxTextureBlendStages   = 8;
2434     caps->MaxSimultaneousTextures = min(GL_LIMITS(fragment_samplers), 8);
2435
2436     caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
2437 }
2438 #undef GLINFO_LOCATION
2439
2440 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
2441 static void state_texfactor_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2442     float col[4];
2443     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2444
2445     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2446      * application provided constants
2447      */
2448     if(device->shader_backend == &arb_program_shader_backend) {
2449         if (use_ps(stateblock)) return;
2450
2451         device = stateblock->wineD3DDevice;
2452         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_TFACTOR] = 1;
2453         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_TFACTOR + 1);
2454     }
2455
2456     D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
2457     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col));
2458     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col)");
2459
2460 }
2461
2462 static void state_arb_specularenable(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2463     float col[4];
2464     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2465
2466     /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2467      * application provided constants
2468      */
2469     if(device->shader_backend == &arb_program_shader_backend) {
2470         if (use_ps(stateblock)) return;
2471
2472         device = stateblock->wineD3DDevice;
2473         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_SPECULAR_ENABLE] = 1;
2474         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_SPECULAR_ENABLE + 1);
2475     }
2476
2477     if(stateblock->renderState[WINED3DRS_SPECULARENABLE]) {
2478         /* The specular color has no alpha */
2479         col[0] = 1.0; col[1] = 1.0;
2480         col[2] = 1.0; col[3] = 0.0;
2481     } else {
2482         col[0] = 0.0; col[1] = 0.0;
2483         col[2] = 0.0; col[3] = 0.0;
2484     }
2485     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col));
2486     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col)");
2487 }
2488
2489 static void set_bumpmat_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2490     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2491     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2492     float mat[2][2];
2493
2494     if (use_ps(stateblock))
2495     {
2496         if(stage != 0 &&
2497            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.bumpmat[stage]) {
2498             /* The pixel shader has to know the bump env matrix. Do a constants update if it isn't scheduled
2499              * anyway
2500              */
2501             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2502                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2503             }
2504         }
2505
2506         if(device->shader_backend == &arb_program_shader_backend) {
2507             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2508             return;
2509         }
2510     } else if(device->shader_backend == &arb_program_shader_backend) {
2511         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_BUMPMAT(stage)] = 1;
2512         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_BUMPMAT(stage) + 1);
2513     }
2514
2515     mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
2516     mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
2517     mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
2518     mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
2519
2520     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0]));
2521     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0])");
2522 }
2523
2524 static void tex_bumpenvlum_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2525     DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2526     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2527     float param[4];
2528
2529     if (use_ps(stateblock))
2530     {
2531         if(stage != 0 &&
2532            ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.luminanceparams[stage]) {
2533             /* The pixel shader has to know the luminance offset. Do a constants update if it
2534              * isn't scheduled anyway
2535              */
2536             if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2537                 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2538             }
2539         }
2540
2541         if(device->shader_backend == &arb_program_shader_backend) {
2542             /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2543             return;
2544         }
2545     } else if(device->shader_backend == &arb_program_shader_backend) {
2546         device->activeContext->pshader_const_dirty[ARB_FFP_CONST_LUMINANCE(stage)] = 1;
2547         device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_LUMINANCE(stage) + 1);
2548     }
2549
2550     param[0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLSCALE]);
2551     param[1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLOFFSET]);
2552     param[2] = 0.0;
2553     param[3] = 0.0;
2554
2555     GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param));
2556     checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param)");
2557 }
2558
2559 static const char *get_argreg(SHADER_BUFFER *buffer, DWORD argnum, unsigned int stage, DWORD arg) {
2560     const char *ret;
2561
2562     if(arg == ARG_UNUSED) return "unused"; /* This is the marker for unused registers */
2563
2564     switch(arg & WINED3DTA_SELECTMASK) {
2565         case WINED3DTA_DIFFUSE:
2566             ret = "fragment.color.primary"; break;
2567
2568         case WINED3DTA_CURRENT:
2569             if(stage == 0) ret = "fragment.color.primary";
2570             else ret = "ret";
2571             break;
2572
2573         case WINED3DTA_TEXTURE:
2574             switch(stage) {
2575                 case 0: ret = "tex0"; break;
2576                 case 1: ret = "tex1"; break;
2577                 case 2: ret = "tex2"; break;
2578                 case 3: ret = "tex3"; break;
2579                 case 4: ret = "tex4"; break;
2580                 case 5: ret = "tex5"; break;
2581                 case 6: ret = "tex6"; break;
2582                 case 7: ret = "tex7"; break;
2583                 default: ret = "unknown texture";
2584             }
2585             break;
2586
2587         case WINED3DTA_TFACTOR:
2588             ret = "tfactor"; break;
2589
2590         case WINED3DTA_SPECULAR:
2591             ret = "fragment.color.secondary"; break;
2592
2593         case WINED3DTA_TEMP:
2594             ret = "tempreg"; break;
2595
2596         case WINED3DTA_CONSTANT:
2597             FIXME("Implement perstage constants\n");
2598             switch(stage) {
2599                 case 0: ret = "const0"; break;
2600                 case 1: ret = "const1"; break;
2601                 case 2: ret = "const2"; break;
2602                 case 3: ret = "const3"; break;
2603                 case 4: ret = "const4"; break;
2604                 case 5: ret = "const5"; break;
2605                 case 6: ret = "const6"; break;
2606                 case 7: ret = "const7"; break;
2607                 default: ret = "unknown constant";
2608             }
2609             break;
2610
2611         default:
2612             return "unknown";
2613     }
2614
2615     if(arg & WINED3DTA_COMPLEMENT) {
2616         shader_addline(buffer, "SUB arg%u, const.x, %s;\n", argnum, ret);
2617         if(argnum == 0) ret = "arg0";
2618         if(argnum == 1) ret = "arg1";
2619         if(argnum == 2) ret = "arg2";
2620     }
2621     if(arg & WINED3DTA_ALPHAREPLICATE) {
2622         shader_addline(buffer, "MOV arg%u, %s.w;\n", argnum, ret);
2623         if(argnum == 0) ret = "arg0";
2624         if(argnum == 1) ret = "arg1";
2625         if(argnum == 2) ret = "arg2";
2626     }
2627     return ret;
2628 }
2629
2630 static void gen_ffp_instr(SHADER_BUFFER *buffer, unsigned int stage, BOOL color, BOOL alpha,
2631                           DWORD dst, DWORD op, DWORD dw_arg0, DWORD dw_arg1, DWORD dw_arg2) {
2632     const char *dstmask, *dstreg, *arg0, *arg1, *arg2;
2633     unsigned int mul = 1;
2634     BOOL mul_final_dest = FALSE;
2635
2636     if(color && alpha) dstmask = "";
2637     else if(color) dstmask = ".xyz";
2638     else dstmask = ".w";
2639
2640     if(dst == tempreg) dstreg = "tempreg";
2641     else dstreg = "ret";
2642
2643     arg0 = get_argreg(buffer, 0, stage, dw_arg0);
2644     arg1 = get_argreg(buffer, 1, stage, dw_arg1);
2645     arg2 = get_argreg(buffer, 2, stage, dw_arg2);
2646
2647     switch(op) {
2648         case WINED3DTOP_DISABLE:
2649             if(stage == 0) shader_addline(buffer, "MOV %s%s, fragment.color.primary;\n", dstreg, dstmask);
2650             break;
2651
2652         case WINED3DTOP_SELECTARG2:
2653             arg1 = arg2;
2654         case WINED3DTOP_SELECTARG1:
2655             shader_addline(buffer, "MOV %s%s, %s;\n", dstreg, dstmask, arg1);
2656             break;
2657
2658         case WINED3DTOP_MODULATE4X:
2659             mul = 2;
2660         case WINED3DTOP_MODULATE2X:
2661             mul *= 2;
2662             if(strcmp(dstreg, "result.color") == 0) {
2663                 dstreg = "ret";
2664                 mul_final_dest = TRUE;
2665             }
2666         case WINED3DTOP_MODULATE:
2667             shader_addline(buffer, "MUL %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2668             break;
2669
2670         case WINED3DTOP_ADDSIGNED2X:
2671             mul = 2;
2672             if(strcmp(dstreg, "result.color") == 0) {
2673                 dstreg = "ret";
2674                 mul_final_dest = TRUE;
2675             }
2676         case WINED3DTOP_ADDSIGNED:
2677             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2678             arg2 = "arg2";
2679         case WINED3DTOP_ADD:
2680             shader_addline(buffer, "ADD_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2681             break;
2682
2683         case WINED3DTOP_SUBTRACT:
2684             shader_addline(buffer, "SUB_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2685             break;
2686
2687         case WINED3DTOP_ADDSMOOTH:
2688             shader_addline(buffer, "SUB arg1, const.x, %s;\n", arg1);
2689             shader_addline(buffer, "MAD_SAT %s%s, arg1, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2690             break;
2691
2692         case WINED3DTOP_BLENDCURRENTALPHA:
2693             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_CURRENT);
2694             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2695             break;
2696         case WINED3DTOP_BLENDFACTORALPHA:
2697             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TFACTOR);
2698             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2699             break;
2700         case WINED3DTOP_BLENDTEXTUREALPHA:
2701             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2702             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2703             break;
2704         case WINED3DTOP_BLENDDIFFUSEALPHA:
2705             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_DIFFUSE);
2706             shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2707             break;
2708
2709         case WINED3DTOP_BLENDTEXTUREALPHAPM:
2710             arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2711             shader_addline(buffer, "SUB arg0.w, const.x, %s.w;\n", arg0);
2712             shader_addline(buffer, "MAD_SAT %s%s, %s, arg0.w, %s;\n", dstreg, dstmask, arg2, arg1);
2713             break;
2714
2715         /* D3DTOP_PREMODULATE ???? */
2716
2717         case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
2718             shader_addline(buffer, "SUB arg0.w, const.x, %s;\n", arg1);
2719             shader_addline(buffer, "MAD_SAT %s%s, arg0.w, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2720             break;
2721         case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
2722             shader_addline(buffer, "MAD_SAT %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg1);
2723             break;
2724         case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
2725             shader_addline(buffer, "SUB arg0, const.x, %s;\n", arg1);
2726             shader_addline(buffer, "MAD_SAT %s%s, arg0, %s, %s.w;\n", dstreg, dstmask, arg2, arg1);
2727             break;
2728         case WINED3DTOP_MODULATECOLOR_ADDALPHA:
2729             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s.w;\n", dstreg, dstmask, arg1, arg2, arg1);
2730             break;
2731
2732         case WINED3DTOP_DOTPRODUCT3:
2733             mul = 4;
2734             if(strcmp(dstreg, "result.color") == 0) {
2735                 dstreg = "ret";
2736                 mul_final_dest = TRUE;
2737             }
2738             shader_addline(buffer, "SUB arg1, %s, const.w;\n", arg1);
2739             shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2740             shader_addline(buffer, "DP3_SAT %s%s, arg1, arg2;\n", dstreg, dstmask);
2741             break;
2742
2743         case WINED3DTOP_MULTIPLYADD:
2744             shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg0);
2745             break;
2746
2747         case WINED3DTOP_LERP:
2748             /* The msdn is not quite right here */
2749             shader_addline(buffer, "LRP %s%s, %s, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2750             break;
2751
2752         case WINED3DTOP_BUMPENVMAP:
2753         case WINED3DTOP_BUMPENVMAPLUMINANCE:
2754             /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
2755             break;
2756
2757         default:
2758             FIXME("Unhandled texture op %08x\n", op);
2759     }
2760
2761     if(mul == 2) {
2762         shader_addline(buffer, "MUL_SAT %s%s, %s, const.y;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2763     } else if(mul == 4) {
2764         shader_addline(buffer, "MUL_SAT %s%s, %s, const.z;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2765     }
2766 }
2767
2768 /* The stateblock is passed for GLINFO_LOCATION */
2769 static GLuint gen_arbfp_ffp_shader(const struct ffp_frag_settings *settings, IWineD3DStateBlockImpl *stateblock)
2770 {
2771     unsigned int stage;
2772     SHADER_BUFFER buffer;
2773     BOOL tex_read[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2774     BOOL bump_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2775     BOOL luminance_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2776     const char *textype;
2777     const char *instr, *sat;
2778     char colorcor_dst[8];
2779     GLuint ret;
2780     DWORD arg0, arg1, arg2;
2781     BOOL tempreg_used = FALSE, tfactor_used = FALSE;
2782     BOOL op_equal;
2783     const char *final_combiner_src = "ret";
2784
2785     /* Find out which textures are read */
2786     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2787         if(settings->op[stage].cop == WINED3DTOP_DISABLE) break;
2788         arg0 = settings->op[stage].carg0 & WINED3DTA_SELECTMASK;
2789         arg1 = settings->op[stage].carg1 & WINED3DTA_SELECTMASK;
2790         arg2 = settings->op[stage].carg2 & WINED3DTA_SELECTMASK;
2791         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2792         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2793         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2794
2795         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) tex_read[stage] = TRUE;
2796         if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHAPM) tex_read[stage] = TRUE;
2797         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP) {
2798             bump_used[stage] = TRUE;
2799             tex_read[stage] = TRUE;
2800         }
2801         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2802             bump_used[stage] = TRUE;
2803             tex_read[stage] = TRUE;
2804             luminance_used[stage] = TRUE;
2805         } else if(settings->op[stage].cop == WINED3DTOP_BLENDFACTORALPHA) {
2806             tfactor_used = TRUE;
2807         }
2808
2809         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2810             tfactor_used = TRUE;
2811         }
2812
2813         if(settings->op[stage].dst == tempreg) tempreg_used = TRUE;
2814         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2815             tempreg_used = TRUE;
2816         }
2817
2818         if(settings->op[stage].aop == WINED3DTOP_DISABLE) continue;
2819         arg0 = settings->op[stage].aarg0 & WINED3DTA_SELECTMASK;
2820         arg1 = settings->op[stage].aarg1 & WINED3DTA_SELECTMASK;
2821         arg2 = settings->op[stage].aarg2 & WINED3DTA_SELECTMASK;
2822         if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2823         if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2824         if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2825
2826         if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2827             tempreg_used = TRUE;
2828         }
2829         if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2830             tfactor_used = TRUE;
2831         }
2832     }
2833
2834     /* Shader header */
2835     shader_buffer_init(&buffer);
2836
2837     shader_addline(&buffer, "!!ARBfp1.0\n");
2838
2839     switch(settings->fog) {
2840         case FOG_OFF:                                                         break;
2841         case FOG_LINEAR: shader_addline(&buffer, "OPTION ARB_fog_linear;\n"); break;
2842         case FOG_EXP:    shader_addline(&buffer, "OPTION ARB_fog_exp;\n");    break;
2843         case FOG_EXP2:   shader_addline(&buffer, "OPTION ARB_fog_exp2;\n");   break;
2844         default: FIXME("Unexpected fog setting %d\n", settings->fog);
2845     }
2846
2847     shader_addline(&buffer, "PARAM const = {1, 2, 4, 0.5};\n");
2848     shader_addline(&buffer, "TEMP TMP;\n");
2849     shader_addline(&buffer, "TEMP ret;\n");
2850     if(tempreg_used || settings->sRGB_write) shader_addline(&buffer, "TEMP tempreg;\n");
2851     shader_addline(&buffer, "TEMP arg0;\n");
2852     shader_addline(&buffer, "TEMP arg1;\n");
2853     shader_addline(&buffer, "TEMP arg2;\n");
2854     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2855         if(!tex_read[stage]) continue;
2856         shader_addline(&buffer, "TEMP tex%u;\n", stage);
2857         if(!bump_used[stage]) continue;
2858         shader_addline(&buffer, "PARAM bumpmat%u = program.env[%u];\n", stage, ARB_FFP_CONST_BUMPMAT(stage));
2859         if(!luminance_used[stage]) continue;
2860         shader_addline(&buffer, "PARAM luminance%u = program.env[%u];\n", stage, ARB_FFP_CONST_LUMINANCE(stage));
2861     }
2862     if(tfactor_used) {
2863         shader_addline(&buffer, "PARAM tfactor = program.env[%u];\n", ARB_FFP_CONST_TFACTOR);
2864     }
2865         shader_addline(&buffer, "PARAM specular_enable = program.env[%u];\n", ARB_FFP_CONST_SPECULAR_ENABLE);
2866
2867     if(settings->sRGB_write) {
2868         shader_addline(&buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
2869                        srgb_mul_low, srgb_mul_low, srgb_mul_low);
2870         shader_addline(&buffer, "PARAM srgb_comparison =  {%f, %f, %f, %f};\n",
2871                        srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
2872         shader_addline(&buffer, "PARAM srgb_pow =  {%f, %f, %f, 1.0};\n",
2873                        srgb_pow, srgb_pow, srgb_pow);
2874         shader_addline(&buffer, "PARAM srgb_mul_hi =  {%f, %f, %f, 1.0};\n",
2875                        srgb_mul_high, srgb_mul_high, srgb_mul_high);
2876         shader_addline(&buffer, "PARAM srgb_sub_hi =  {%f, %f, %f, 0.0};\n",
2877                        srgb_sub_high, srgb_sub_high, srgb_sub_high);
2878     }
2879
2880     /* Generate texture sampling instructions) */
2881     for(stage = 0; stage < MAX_TEXTURES && settings->op[stage].cop != WINED3DTOP_DISABLE; stage++) {
2882         if(!tex_read[stage]) continue;
2883
2884         switch(settings->op[stage].tex_type) {
2885             case tex_1d:                    textype = "1D";     break;
2886             case tex_2d:                    textype = "2D";     break;
2887             case tex_3d:                    textype = "3D";     break;
2888             case tex_cube:                  textype = "CUBE";   break;
2889             case tex_rect:                  textype = "RECT";   break;
2890             default: textype = "unexpected_textype";   break;
2891         }
2892
2893         if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP ||
2894            settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2895             sat = "";
2896         } else {
2897             sat = "_SAT";
2898         }
2899
2900         if(settings->op[stage].projected == proj_none) {
2901             instr = "TEX";
2902         } else if(settings->op[stage].projected == proj_count4 ||
2903                   settings->op[stage].projected == proj_count3) {
2904             instr = "TXP";
2905         } else {
2906             FIXME("Unexpected projection mode %d\n", settings->op[stage].projected);
2907             instr = "TXP";
2908         }
2909
2910         if(stage > 0 &&
2911            (settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
2912             settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
2913             shader_addline(&buffer, "SWZ arg1, bumpmat%u, x, z, 0, 0;\n", stage - 1);
2914             shader_addline(&buffer, "DP3 ret.x, arg1, tex%u;\n", stage - 1);
2915             shader_addline(&buffer, "SWZ arg1, bumpmat%u, y, w, 0, 0;\n", stage - 1);
2916             shader_addline(&buffer, "DP3 ret.y, arg1, tex%u;\n", stage - 1);
2917
2918             /* with projective textures, texbem only divides the static texture coord, not the displacement,
2919              * so multiply the displacement with the dividing parameter before passing it to TXP
2920              */
2921             if (settings->op[stage].projected != proj_none) {
2922                 if(settings->op[stage].projected == proj_count4) {
2923                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].w;\n", stage);
2924                     shader_addline(&buffer, "MUL ret.xyz, ret, fragment.texcoord[%u].w, fragment.texcoord[%u];\n", stage, stage);
2925                 } else {
2926                     shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].z;\n", stage);
2927                     shader_addline(&buffer, "MAD ret.xyz, ret, fragment.texcoord[%u].z, fragment.texcoord[%u];\n", stage, stage);
2928                 }
2929             } else {
2930                 shader_addline(&buffer, "ADD ret, ret, fragment.texcoord[%u];\n", stage);
2931             }
2932
2933             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2934                            instr, sat, stage, stage, textype);
2935             if(settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2936                 shader_addline(&buffer, "MAD_SAT ret.x, tex%u.z, luminance%u.x, luminance%u.y;\n",
2937                                stage - 1, stage - 1, stage - 1);
2938                 shader_addline(&buffer, "MUL tex%u, tex%u, ret.x;\n", stage, stage);
2939             }
2940         } else if(settings->op[stage].projected == proj_count3) {
2941             shader_addline(&buffer, "MOV ret, fragment.texcoord[%u];\n", stage);
2942             shader_addline(&buffer, "MOV ret.w, ret.z;\n");
2943             shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2944                             instr, sat, stage, stage, textype);
2945         } else {
2946             shader_addline(&buffer, "%s%s tex%u, fragment.texcoord[%u], texture[%u], %s;\n",
2947                             instr, sat, stage, stage, stage, textype);
2948         }
2949
2950         sprintf(colorcor_dst, "tex%u", stage);
2951         gen_color_correction(&buffer, colorcor_dst, WINED3DSP_WRITEMASK_ALL, "const.x", "const.y",
2952                 settings->op[stage].color_fixup);
2953     }
2954
2955     /* Generate the main shader */
2956     for(stage = 0; stage < MAX_TEXTURES; stage++) {
2957         if(settings->op[stage].cop == WINED3DTOP_DISABLE) {
2958             if(stage == 0) {
2959                 final_combiner_src = "fragment.color.primary";
2960             }
2961             break;
2962         }
2963
2964         if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2965            settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2966             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg1;
2967         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2968                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2969             op_equal = settings->op[stage].carg1 == settings->op[stage].aarg2;
2970         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2971                   settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2972             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg1;
2973         } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2974                   settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2975             op_equal = settings->op[stage].carg2 == settings->op[stage].aarg2;
2976         } else {
2977             op_equal = settings->op[stage].aop   == settings->op[stage].cop &&
2978                        settings->op[stage].carg0 == settings->op[stage].aarg0 &&
2979                        settings->op[stage].carg1 == settings->op[stage].aarg1 &&
2980                        settings->op[stage].carg2 == settings->op[stage].aarg2;
2981         }
2982
2983         if(settings->op[stage].aop == WINED3DTOP_DISABLE) {
2984             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2985                           settings->op[stage].cop, settings->op[stage].carg0,
2986                           settings->op[stage].carg1, settings->op[stage].carg2);
2987             if(stage == 0) {
2988                 shader_addline(&buffer, "MOV ret.w, fragment.color.primary.w;\n");
2989             }
2990         } else if(op_equal) {
2991             gen_ffp_instr(&buffer, stage, TRUE, TRUE, settings->op[stage].dst,
2992                           settings->op[stage].cop, settings->op[stage].carg0,
2993                           settings->op[stage].carg1, settings->op[stage].carg2);
2994         } else {
2995             gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2996                           settings->op[stage].cop, settings->op[stage].carg0,
2997                           settings->op[stage].carg1, settings->op[stage].carg2);
2998             gen_ffp_instr(&buffer, stage, FALSE, TRUE, settings->op[stage].dst,
2999                           settings->op[stage].aop, settings->op[stage].aarg0,
3000                           settings->op[stage].aarg1, settings->op[stage].aarg2);
3001         }
3002     }
3003
3004     if(settings->sRGB_write) {
3005         shader_addline(&buffer, "MAD ret, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
3006         arbfp_add_sRGB_correction(&buffer, "ret", "arg0", "arg1", "arg2", "tempreg");
3007         shader_addline(&buffer, "MOV result.color.w, ret.w;\n");
3008     } else {
3009         shader_addline(&buffer, "MAD result.color, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
3010     }
3011
3012     /* Footer */
3013     shader_addline(&buffer, "END\n");
3014
3015     /* Generate the shader */
3016     GL_EXTCALL(glGenProgramsARB(1, &ret));
3017     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, ret));
3018     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3019
3020     if (glGetError() == GL_INVALID_OPERATION) {
3021         GLint pos;
3022         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3023         FIXME("Fragment program error at position %d: %s\n", pos,
3024               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3025     }
3026     shader_buffer_free(&buffer);
3027     return ret;
3028 }
3029
3030 static void fragment_prog_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3031     IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
3032     struct shader_arb_priv *priv = device->fragment_priv;
3033     BOOL use_pshader = use_ps(stateblock);
3034     BOOL use_vshader = use_vs(stateblock);
3035     struct ffp_frag_settings settings;
3036     const struct arbfp_ffp_desc *desc;
3037     unsigned int i;
3038
3039     TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3040
3041     if(isStateDirty(context, STATE_RENDER(WINED3DRS_FOGENABLE))) {
3042         if(!use_pshader && device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3043             /* Reload fixed function constants since they collide with the pixel shader constants */
3044             for(i = 0; i < MAX_TEXTURES; i++) {
3045                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3046             }
3047             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3048             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3049         } else if(use_pshader && !isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3050             device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3051         }
3052         return;
3053     }
3054
3055     if(!use_pshader) {
3056         /* Find or create a shader implementing the fixed function pipeline settings, then activate it */
3057         gen_ffp_frag_op(stateblock, &settings, FALSE);
3058         desc = (const struct arbfp_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
3059         if(!desc) {
3060             struct arbfp_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
3061             if (!new_desc)
3062             {
3063                 ERR("Out of memory\n");
3064                 return;
3065             }
3066             new_desc->num_textures_used = 0;
3067             for(i = 0; i < GL_LIMITS(texture_stages); i++) {
3068                 if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
3069                 new_desc->num_textures_used = i;
3070             }
3071
3072             memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
3073             new_desc->shader = gen_arbfp_ffp_shader(&settings, stateblock);
3074             add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
3075             TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
3076             desc = new_desc;
3077         }
3078
3079         /* Now activate the replacement program. GL_FRAGMENT_PROGRAM_ARB is already active(however, note the
3080          * comment above the shader_select call below). If e.g. GLSL is active, the shader_select call will
3081          * deactivate it.
3082          */
3083         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader));
3084         checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader)");
3085         priv->current_fprogram_id = desc->shader;
3086
3087         if(device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3088             /* Reload fixed function constants since they collide with the pixel shader constants */
3089             for(i = 0; i < MAX_TEXTURES; i++) {
3090                 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3091             }
3092             state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3093             state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3094         }
3095         context->last_was_pshader = FALSE;
3096     } else {
3097         context->last_was_pshader = TRUE;
3098     }
3099
3100     /* Finally, select the shader. If a pixel shader is used, it will be set and enabled by the shader backend.
3101      * If this shader backend is arbfp(most likely), then it will simply overwrite the last fixed function replace-
3102      * ment shader. If the shader backend is not ARB, it currently is important that the opengl implementation
3103      * type overwrites GL_ARB_fragment_program. This is currently the case with GLSL. If we really want to use
3104      * atifs or nvrc pixel shaders with arb fragment programs we'd have to disable GL_FRAGMENT_PROGRAM_ARB here
3105      *
3106      * Don't call shader_select if the vertex shader is dirty, because it will be called later on by the vertex
3107      * shader handler
3108      */
3109     if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3110         device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3111
3112         if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && (use_vshader || use_pshader)) {
3113             device->StateTable[STATE_VERTEXSHADERCONSTANT].apply(STATE_VERTEXSHADERCONSTANT, stateblock, context);
3114         }
3115     }
3116     if(use_pshader) {
3117         device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
3118     }
3119 }
3120
3121 /* We can't link the fog states to the fragment state directly since the vertex pipeline links them
3122  * to FOGENABLE. A different linking in different pipeline parts can't be expressed in the combined
3123  * state table, so we need to handle that with a forwarding function. The other invisible side effect
3124  * is that changing the fog start and fog end(which links to FOGENABLE in vertex) results in the
3125  * fragment_prog_arbfp function being called because FOGENABLE is dirty, which calls this function here
3126  */
3127 static void state_arbfp_fog(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3128     enum fogsource new_source;
3129
3130     TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3131
3132     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3133         fragment_prog_arbfp(state, stateblock, context);
3134     }
3135
3136     if(!stateblock->renderState[WINED3DRS_FOGENABLE]) return;
3137
3138     if(stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
3139         if(use_vs(stateblock)) {
3140             new_source = FOGSOURCE_VS;
3141         } else {
3142             if(stateblock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || context->last_was_rhw) {
3143                 new_source = FOGSOURCE_COORD;
3144             } else {
3145                 new_source = FOGSOURCE_FFP;
3146             }
3147         }
3148     } else {
3149         new_source = FOGSOURCE_FFP;
3150     }
3151     if(new_source != context->fog_source) {
3152         context->fog_source = new_source;
3153         state_fogstartend(STATE_RENDER(WINED3DRS_FOGSTART), stateblock, context);
3154     }
3155 }
3156
3157 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3158     if(!isStateDirty(context, STATE_PIXELSHADER)) {
3159         fragment_prog_arbfp(state, stateblock, context);
3160     }
3161 }
3162
3163 #undef GLINFO_LOCATION
3164
3165 static const struct StateEntryTemplate arbfp_fragmentstate_template[] = {
3166     {STATE_RENDER(WINED3DRS_TEXTUREFACTOR),               { STATE_RENDER(WINED3DRS_TEXTUREFACTOR),              state_texfactor_arbfp   }, 0                               },
3167     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3168     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3169     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3170     {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3171     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3172     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3173     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3174     {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3175     {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3176     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3177     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3178     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3179     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3180     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3181     {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3182     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3183     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3184     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3185     {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3186     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3187     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3188     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3189     {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3190     {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3191     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3192     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3193     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3194     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3195     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3196     {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3197     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3198     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3199     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3200     {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3201     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3202     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3203     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3204     {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3205     {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3206     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3207     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3208     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3209     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3210     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3211     {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3212     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3213     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3214     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3215     {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3216     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3217     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3218     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3219     {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3220     {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3221     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3222     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3223     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3224     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3225     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3226     {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3227     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3228     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3229     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3230     {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3231     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3232     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3233     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3234     {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3235     {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3236     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3237     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3238     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3239     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3240     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3241     {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3242     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3243     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3244     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3245     {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3246     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3247     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3248     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3249     {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3250     {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3251     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3252     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3253     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3254     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3255     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3256     {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3257     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3258     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3259     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3260     {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3261     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3262     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3263     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3264     {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3265     {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3266     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3267     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3268     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3269     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3270     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3271     {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3272     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3273     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3274     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3275     {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3276     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP),           { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3277     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3278     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3279     {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3280     {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG),         { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3281     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3282     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3283     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3284     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11),      { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00),     set_bumpmat_arbfp       }, 0                               },
3285     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),     { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3286     {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLOFFSET),    { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE),    tex_bumpenvlum_arbfp    }, 0                               },
3287     {STATE_SAMPLER(0),                                    { STATE_SAMPLER(0),                                   sampler_texdim          }, 0                               },
3288     {STATE_SAMPLER(1),                                    { STATE_SAMPLER(1),                                   sampler_texdim          }, 0                               },
3289     {STATE_SAMPLER(2),                                    { STATE_SAMPLER(2),                                   sampler_texdim          }, 0                               },
3290     {STATE_SAMPLER(3),                                    { STATE_SAMPLER(3),                                   sampler_texdim          }, 0                               },
3291     {STATE_SAMPLER(4),                                    { STATE_SAMPLER(4),                                   sampler_texdim          }, 0                               },
3292     {STATE_SAMPLER(5),                                    { STATE_SAMPLER(5),                                   sampler_texdim          }, 0                               },
3293     {STATE_SAMPLER(6),                                    { STATE_SAMPLER(6),                                   sampler_texdim          }, 0                               },
3294     {STATE_SAMPLER(7),                                    { STATE_SAMPLER(7),                                   sampler_texdim          }, 0                               },
3295     {STATE_PIXELSHADER,                                   { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3296     {STATE_RENDER(WINED3DRS_FOGENABLE),                   { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3297     {STATE_RENDER(WINED3DRS_FOGTABLEMODE),                { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3298     {STATE_RENDER(WINED3DRS_FOGVERTEXMODE),               { STATE_RENDER(WINED3DRS_FOGENABLE),                  state_arbfp_fog         }, 0                               },
3299     {STATE_RENDER(WINED3DRS_FOGSTART),                    { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, 0                               },
3300     {STATE_RENDER(WINED3DRS_FOGEND),                      { STATE_RENDER(WINED3DRS_FOGSTART),                   state_fogstartend       }, 0                               },
3301     {STATE_RENDER(WINED3DRS_SRGBWRITEENABLE),             { STATE_PIXELSHADER,                                  fragment_prog_arbfp     }, 0                               },
3302     {STATE_RENDER(WINED3DRS_FOGCOLOR),                    { STATE_RENDER(WINED3DRS_FOGCOLOR),                   state_fogcolor          }, 0                               },
3303     {STATE_RENDER(WINED3DRS_FOGDENSITY),                  { STATE_RENDER(WINED3DRS_FOGDENSITY),                 state_fogdensity        }, 0                               },
3304     {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3305     {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3306     {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3307     {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3308     {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3309     {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3310     {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3311     {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform      }, 0                               },
3312     {STATE_RENDER(WINED3DRS_SPECULARENABLE),              { STATE_RENDER(WINED3DRS_SPECULARENABLE),             state_arb_specularenable}, 0                               },
3313     {0 /* Terminate */,                                   { 0,                                                  0                       }, 0                               },
3314 };
3315
3316 const struct fragment_pipeline arbfp_fragment_pipeline = {
3317     arbfp_enable,
3318     arbfp_get_caps,
3319     arbfp_alloc,
3320     arbfp_free,
3321     shader_arb_color_fixup_supported,
3322     arbfp_fragmentstate_template,
3323     TRUE /* We can disable projected textures */
3324 };
3325
3326 #define GLINFO_LOCATION device->adapter->gl_info
3327
3328 struct arbfp_blit_priv {
3329     GLenum yuy2_rect_shader, yuy2_2d_shader;
3330     GLenum uyvy_rect_shader, uyvy_2d_shader;
3331     GLenum yv12_rect_shader, yv12_2d_shader;
3332 };
3333
3334 static HRESULT arbfp_blit_alloc(IWineD3DDevice *iface) {
3335     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3336     device->blit_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct arbfp_blit_priv));
3337     if(!device->blit_priv) {
3338         ERR("Out of memory\n");
3339         return E_OUTOFMEMORY;
3340     }
3341     return WINED3D_OK;
3342 }
3343 static void arbfp_blit_free(IWineD3DDevice *iface) {
3344     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3345     struct arbfp_blit_priv *priv = device->blit_priv;
3346
3347     ENTER_GL();
3348     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_rect_shader));
3349     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_2d_shader));
3350     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_rect_shader));
3351     GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_2d_shader));
3352     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_rect_shader));
3353     GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_2d_shader));
3354     checkGLcall("Delete yuv programs\n");
3355     LEAVE_GL();
3356 }
3357
3358 static BOOL gen_planar_yuv_read(SHADER_BUFFER *buffer, enum yuv_fixup yuv_fixup, GLenum textype, char *luminance)
3359 {
3360     char chroma;
3361     const char *tex, *texinstr;
3362
3363     if (yuv_fixup == YUV_FIXUP_UYVY) {
3364         chroma = 'x';
3365         *luminance = 'w';
3366     } else {
3367         chroma = 'w';
3368         *luminance = 'x';
3369     }
3370     switch(textype) {
3371         case GL_TEXTURE_2D:             tex = "2D";     texinstr = "TXP"; break;
3372         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   texinstr = "TEX"; break;
3373         default:
3374             /* This is more tricky than just replacing the texture type - we have to navigate
3375              * properly in the texture to find the correct chroma values
3376              */
3377             FIXME("Implement yuv correction for non-2d, non-rect textures\n");
3378             return FALSE;
3379     }
3380
3381     /* First we have to read the chroma values. This means we need at least two pixels(no filtering),
3382      * or 4 pixels(with filtering). To get the unmodified chromas, we have to rid ourselves of the
3383      * filtering when we sample the texture.
3384      *
3385      * These are the rules for reading the chroma:
3386      *
3387      * Even pixel: Cr
3388      * Even pixel: U
3389      * Odd pixel: V
3390      *
3391      * So we have to get the sampling x position in non-normalized coordinates in integers
3392      */
3393     if(textype != GL_TEXTURE_RECTANGLE_ARB) {
3394         shader_addline(buffer, "MUL texcrd.xy, fragment.texcoord[0], size.x;\n");
3395         shader_addline(buffer, "MOV texcrd.w, size.x;\n");
3396     } else {
3397         shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3398     }
3399     /* We must not allow filtering between pixel x and x+1, this would mix U and V
3400      * Vertical filtering is ok. However, bear in mind that the pixel center is at
3401      * 0.5, so add 0.5.
3402      */
3403     shader_addline(buffer, "FLR texcrd.x, texcrd.x;\n");
3404     shader_addline(buffer, "ADD texcrd.x, texcrd.x, coef.y;\n");
3405
3406     /* Divide the x coordinate by 0.5 and get the fraction. This gives 0.25 and 0.75 for the
3407      * even and odd pixels respectively
3408      */
3409     shader_addline(buffer, "MUL texcrd2, texcrd, coef.y;\n");
3410     shader_addline(buffer, "FRC texcrd2, texcrd2;\n");
3411
3412     /* Sample Pixel 1 */
3413     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3414
3415     /* Put the value into either of the chroma values */
3416     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3417     shader_addline(buffer, "MUL chroma.x, luminance.%c, temp.x;\n", chroma);
3418     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3419     shader_addline(buffer, "MUL chroma.y, luminance.%c, temp.x;\n", chroma);
3420
3421     /* Sample pixel 2. If we read an even pixel(SLT above returned 1), sample
3422      * the pixel right to the current one. Otherwise, sample the left pixel.
3423      * Bias and scale the SLT result to -1;1 and add it to the texcrd.x.
3424      */
3425     shader_addline(buffer, "MAD temp.x, temp.x, coef.z, -coef.x;\n");
3426     shader_addline(buffer, "ADD texcrd.x, texcrd, temp.x;\n");
3427     shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3428
3429     /* Put the value into the other chroma */
3430     shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3431     shader_addline(buffer, "MAD chroma.y, luminance.%c, temp.x, chroma.y;\n", chroma);
3432     shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3433     shader_addline(buffer, "MAD chroma.x, luminance.%c, temp.x, chroma.x;\n", chroma);
3434
3435     /* TODO: If filtering is enabled, sample a 2nd pair of pixels left or right of
3436      * the current one and lerp the two U and V values
3437      */
3438
3439     /* This gives the correctly filtered luminance value */
3440     shader_addline(buffer, "TEX luminance, fragment.texcoord[0], texture[0], %s;\n", tex);
3441
3442     return TRUE;
3443 }
3444
3445 static BOOL gen_yv12_read(SHADER_BUFFER *buffer, GLenum textype, char *luminance)
3446 {
3447     const char *tex;
3448
3449     switch(textype) {
3450         case GL_TEXTURE_2D:             tex = "2D";     break;
3451         case GL_TEXTURE_RECTANGLE_ARB:  tex = "RECT";   break;
3452         default:
3453             FIXME("Implement yv12 correction for non-2d, non-rect textures\n");
3454             return FALSE;
3455     }
3456
3457     /* YV12 surfaces contain a WxH sized luminance plane, followed by a (W/2)x(H/2)
3458      * V and a (W/2)x(H/2) U plane, each with 8 bit per pixel. So the effective
3459      * bitdepth is 12 bits per pixel. Since the U and V planes have only half the
3460      * pitch of the luminance plane, the packing into the gl texture is a bit
3461      * unfortunate. If the whole texture is interpreted as luminance data it looks
3462      * approximately like this:
3463      *
3464      *        +----------------------------------+----
3465      *        |                                  |
3466      *        |                                  |
3467      *        |                                  |
3468      *        |                                  |
3469      *        |                                  |   2
3470      *        |            LUMINANCE             |   -
3471      *        |                                  |   3
3472      *        |                                  |
3473      *        |                                  |
3474      *        |                                  |
3475      *        |                                  |
3476      *        +----------------+-----------------+----
3477      *        |                |                 |
3478      *        |  U even rows   |  U odd rows     |
3479      *        |                |                 |   1
3480      *        +----------------+------------------   -
3481      *        |                |                 |   3
3482      *        |  V even rows   |  V odd rows     |
3483      *        |                |                 |
3484      *        +----------------+-----------------+----
3485      *        |                |                 |
3486      *        |     0.5        |       0.5       |
3487      *
3488      * So it appears as if there are 4 chroma images, but in fact the odd rows
3489      * in the chroma images are in the same row as the even ones. So its is
3490      * kinda tricky to read
3491      *
3492      * When reading from rectangle textures, keep in mind that the input y coordinates
3493      * go from 0 to d3d_height, whereas the opengl texture height is 1.5 * d3d_height
3494      */
3495     shader_addline(buffer, "PARAM yv12_coef = {%f, %f, %f, %f};\n",
3496                    2.0 / 3.0, 1.0 / 6.0, (2.0 / 3.0) + (1.0 / 6.0), 1.0 / 3.0);
3497
3498     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3499     /* the chroma planes have only half the width */
3500     shader_addline(buffer, "MUL texcrd.x, texcrd.x, coef.y;\n");
3501
3502     /* The first value is between 2/3 and 5/6th of the texture's height, so scale+bias
3503      * the coordinate. Also read the right side of the image when reading odd lines
3504      *
3505      * Don't forget to clamp the y values in into the range, otherwise we'll get filtering
3506      * bleeding
3507      */
3508     if(textype == GL_TEXTURE_2D) {
3509
3510         shader_addline(buffer, "RCP chroma.w, size.y;\n");
3511
3512         shader_addline(buffer, "MUL texcrd2.y, texcrd.y, size.y;\n");
3513
3514         shader_addline(buffer, "FLR texcrd2.y, texcrd2.y;\n");
3515         shader_addline(buffer, "MAD texcrd.y, texcrd.y, yv12_coef.y, yv12_coef.x;\n");
3516
3517         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3518         shader_addline(buffer, "ADD texcrd2.x, texcrd2.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3519         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3520         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3521         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3522
3523         /* clamp, keep the half pixel origin in mind */
3524         shader_addline(buffer, "MAD temp.y, coef.y, chroma.w, yv12_coef.x;\n");
3525         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3526         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.z;\n");
3527         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3528     } else {
3529         /* Read from [size - size+size/4] */
3530         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3531         shader_addline(buffer, "MAD texcrd.y, texcrd.y, coef.w, size.y;\n");
3532
3533         /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3534         shader_addline(buffer, "ADD texcrd2.x, texcrd.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3535         shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3536         shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3537         shader_addline(buffer, "MUL texcrd2.x, texcrd2.x, size.x;\n");
3538         shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3539
3540         /* Make sure to read exactly from the pixel center */
3541         shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3542         shader_addline(buffer, "ADD texcrd.y, texcrd.y, coef.y;\n");
3543
3544         /* Clamp */
3545         shader_addline(buffer, "MAD temp.y, size.y, coef.w, size.y;\n");
3546         shader_addline(buffer, "ADD temp.y, temp.y, -coef.y;\n");
3547         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3548         shader_addline(buffer, "ADD temp.y, size.y, -coef.y;\n");
3549         shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3550     }
3551     /* Read the texture, put the result into the output register */
3552     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3553     shader_addline(buffer, "MOV chroma.x, temp.w;\n");
3554
3555     /* The other chroma value is 1/6th of the texture lower, from 5/6th to 6/6th
3556      * No need to clamp because we're just reusing the already clamped value from above
3557      */
3558     if(textype == GL_TEXTURE_2D) {
3559         shader_addline(buffer, "ADD texcrd.y, texcrd.y, yv12_coef.y;\n");
3560     } else {
3561         shader_addline(buffer, "MAD texcrd.y, size.y, coef.w, texcrd.y;\n");
3562     }
3563     shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3564     shader_addline(buffer, "MOV chroma.y, temp.w;\n");
3565
3566     /* Sample the luminance value. It is in the top 2/3rd of the texture, so scale the y coordinate.
3567      * Clamp the y coordinate to prevent the chroma values from bleeding into the sampled luminance
3568      * values due to filtering
3569      */
3570     shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3571     if(textype == GL_TEXTURE_2D) {
3572         /* Multiply the y coordinate by 2/3 and clamp it */
3573         shader_addline(buffer, "MUL texcrd.y, texcrd.y, yv12_coef.x;\n");
3574         shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.x;\n");
3575         shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3576         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3577     } else {
3578         /* Reading from texture_rectangles is pretty straightforward, just use the unmodified
3579          * texture coordinate. It is still a good idea to clamp it though, since the opengl texture
3580          * is bigger
3581          */
3582         shader_addline(buffer, "ADD temp.x, size.y, -coef.y;\n");
3583         shader_addline(buffer, "MIN texcrd.y, texcrd.y, size.x;\n");
3584         shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3585     }
3586     *luminance = 'a';
3587
3588     return TRUE;
3589 }
3590
3591 static GLuint gen_yuv_shader(IWineD3DDeviceImpl *device, enum yuv_fixup yuv_fixup, GLenum textype)
3592 {
3593     GLenum shader;
3594     SHADER_BUFFER buffer;
3595     char luminance_component;
3596     struct arbfp_blit_priv *priv = device->blit_priv;
3597
3598     /* Shader header */
3599     shader_buffer_init(&buffer);
3600
3601     ENTER_GL();
3602     GL_EXTCALL(glGenProgramsARB(1, &shader));
3603     checkGLcall("GL_EXTCALL(glGenProgramsARB(1, &shader))");
3604     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3605     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3606     LEAVE_GL();
3607     if(!shader) {
3608         shader_buffer_free(&buffer);
3609         return 0;
3610     }
3611
3612     /* The YUY2 and UYVY formats contain two pixels packed into a 32 bit macropixel,
3613      * giving effectively 16 bit per pixel. The color consists of a luminance(Y) and
3614      * two chroma(U and V) values. Each macropixel has two luminance values, one for
3615      * each single pixel it contains, and one U and one V value shared between both
3616      * pixels.
3617      *
3618      * The data is loaded into an A8L8 texture. With YUY2, the luminance component
3619      * contains the luminance and alpha the chroma. With UYVY it is vice versa. Thus
3620      * take the format into account when generating the read swizzles
3621      *
3622      * Reading the Y value is straightforward - just sample the texture. The hardware
3623      * takes care of filtering in the horizontal and vertical direction.
3624      *
3625      * Reading the U and V values is harder. We have to avoid filtering horizontally,
3626      * because that would mix the U and V values of one pixel or two adjacent pixels.
3627      * Thus floor the texture coordinate and add 0.5 to get an unfiltered read,
3628      * regardless of the filtering setting. Vertical filtering works automatically
3629      * though - the U and V values of two rows are mixed nicely.
3630      *
3631      * Appart of avoiding filtering issues, the code has to know which value it just
3632      * read, and where it can find the other one. To determine this, it checks if
3633      * it sampled an even or odd pixel, and shifts the 2nd read accordingly.
3634      *
3635      * Handling horizontal filtering of U and V values requires reading a 2nd pair
3636      * of pixels, extracting U and V and mixing them. This is not implemented yet.
3637      *
3638      * An alternative implementation idea is to load the texture as A8R8G8B8 texture,
3639      * with width / 2. This way one read gives all 3 values, finding U and V is easy
3640      * in an unfiltered situation. Finding the luminance on the other hand requires
3641      * finding out if it is an odd or even pixel. The real drawback of this approach
3642      * is filtering. This would have to be emulated completely in the shader, reading
3643      * up two 2 packed pixels in up to 2 rows and interpolating both horizontally and
3644      * vertically. Beyond that it would require adjustments to the texture handling
3645      * code to deal with the width scaling
3646      */
3647     shader_addline(&buffer, "!!ARBfp1.0\n");
3648     shader_addline(&buffer, "TEMP luminance;\n");
3649     shader_addline(&buffer, "TEMP temp;\n");
3650     shader_addline(&buffer, "TEMP chroma;\n");
3651     shader_addline(&buffer, "TEMP texcrd;\n");
3652     shader_addline(&buffer, "TEMP texcrd2;\n");
3653     shader_addline(&buffer, "PARAM coef = {1.0, 0.5, 2.0, 0.25};\n");
3654     shader_addline(&buffer, "PARAM yuv_coef = {1.403, 0.344, 0.714, 1.770};\n");
3655     shader_addline(&buffer, "PARAM size = program.local[0];\n");
3656
3657     switch (yuv_fixup)
3658     {
3659         case YUV_FIXUP_UYVY:
3660         case YUV_FIXUP_YUY2:
3661             if (!gen_planar_yuv_read(&buffer, yuv_fixup, textype, &luminance_component))
3662             {
3663                 shader_buffer_free(&buffer);
3664                 return 0;
3665             }
3666             break;
3667
3668         case YUV_FIXUP_YV12:
3669             if (!gen_yv12_read(&buffer, textype, &luminance_component))
3670             {
3671                 shader_buffer_free(&buffer);
3672                 return 0;
3673             }
3674             break;
3675
3676         default:
3677             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3678             shader_buffer_free(&buffer);
3679             return 0;
3680     }
3681
3682     /* Calculate the final result. Formula is taken from
3683      * http://www.fourcc.org/fccyvrgb.php. Note that the chroma
3684      * ranges from -0.5 to 0.5
3685      */
3686     shader_addline(&buffer, "SUB chroma.xy, chroma, coef.y;\n");
3687
3688     shader_addline(&buffer, "MAD result.color.x, chroma.x, yuv_coef.x, luminance.%c;\n", luminance_component);
3689     shader_addline(&buffer, "MAD temp.x, -chroma.y, yuv_coef.y, luminance.%c;\n", luminance_component);
3690     shader_addline(&buffer, "MAD result.color.y, -chroma.x, yuv_coef.z, temp.x;\n");
3691     shader_addline(&buffer, "MAD result.color.z, chroma.y, yuv_coef.w, luminance.%c;\n", luminance_component);
3692     shader_addline(&buffer, "END\n");
3693
3694     ENTER_GL();
3695     GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3696
3697     if (glGetError() == GL_INVALID_OPERATION) {
3698         GLint pos;
3699         glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3700         FIXME("Fragment program error at position %d: %s\n", pos,
3701               debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3702     }
3703     shader_buffer_free(&buffer);
3704     LEAVE_GL();
3705
3706     switch (yuv_fixup)
3707     {
3708         case YUV_FIXUP_YUY2:
3709             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yuy2_rect_shader = shader;
3710             else priv->yuy2_2d_shader = shader;
3711             break;
3712
3713         case YUV_FIXUP_UYVY:
3714             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->uyvy_rect_shader = shader;
3715             else priv->uyvy_2d_shader = shader;
3716             break;
3717
3718         case YUV_FIXUP_YV12:
3719             if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yv12_rect_shader = shader;
3720             else priv->yv12_2d_shader = shader;
3721             break;
3722     }
3723
3724     return shader;
3725 }
3726
3727 static HRESULT arbfp_blit_set(IWineD3DDevice *iface, const struct GlPixelFormatDesc *format_desc,
3728         GLenum textype, UINT width, UINT height)
3729 {
3730     GLenum shader;
3731     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3732     float size[4] = {width, height, 1, 1};
3733     struct arbfp_blit_priv *priv = device->blit_priv;
3734     enum yuv_fixup yuv_fixup;
3735
3736     if (!is_yuv_fixup(format_desc->color_fixup))
3737     {
3738         TRACE("Fixup:\n");
3739         dump_color_fixup_desc(format_desc->color_fixup);
3740         /* Don't bother setting up a shader for unconverted formats */
3741         ENTER_GL();
3742         glEnable(textype);
3743         checkGLcall("glEnable(textype)");
3744         LEAVE_GL();
3745         return WINED3D_OK;
3746     }
3747
3748     yuv_fixup = get_yuv_fixup(format_desc->color_fixup);
3749
3750     switch(yuv_fixup)
3751     {
3752         case YUV_FIXUP_YUY2:
3753             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yuy2_rect_shader : priv->yuy2_2d_shader;
3754             break;
3755
3756         case YUV_FIXUP_UYVY:
3757             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->uyvy_rect_shader : priv->uyvy_2d_shader;
3758             break;
3759
3760         case YUV_FIXUP_YV12:
3761             shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yv12_rect_shader : priv->yv12_2d_shader;
3762             break;
3763
3764         default:
3765             FIXME("Unsupported YUV fixup %#x, not setting a shader\n", yuv_fixup);
3766             ENTER_GL();
3767             glEnable(textype);
3768             checkGLcall("glEnable(textype)");
3769             LEAVE_GL();
3770             return E_NOTIMPL;
3771     }
3772
3773     if (!shader) shader = gen_yuv_shader(device, yuv_fixup, textype);
3774
3775     ENTER_GL();
3776     glEnable(GL_FRAGMENT_PROGRAM_ARB);
3777     checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
3778     GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3779     checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3780     GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, 0, size));
3781     checkGLcall("glProgramLocalParameter4fvARB");
3782     LEAVE_GL();
3783
3784     return WINED3D_OK;
3785 }
3786
3787 static void arbfp_blit_unset(IWineD3DDevice *iface) {
3788     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3789
3790     ENTER_GL();
3791     glDisable(GL_FRAGMENT_PROGRAM_ARB);
3792     checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
3793     glDisable(GL_TEXTURE_2D);
3794     checkGLcall("glDisable(GL_TEXTURE_2D)");
3795     if(GL_SUPPORT(ARB_TEXTURE_CUBE_MAP)) {
3796         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
3797         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
3798     }
3799     if(GL_SUPPORT(ARB_TEXTURE_RECTANGLE)) {
3800         glDisable(GL_TEXTURE_RECTANGLE_ARB);
3801         checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
3802     }
3803     LEAVE_GL();
3804 }
3805
3806 static BOOL arbfp_blit_color_fixup_supported(struct color_fixup_desc fixup)
3807 {
3808     enum yuv_fixup yuv_fixup;
3809
3810     if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
3811     {
3812         TRACE("Checking support for fixup:\n");
3813         dump_color_fixup_desc(fixup);
3814     }
3815
3816     if (is_identity_fixup(fixup))
3817     {
3818         TRACE("[OK]\n");
3819         return TRUE;
3820     }
3821
3822     /* We only support YUV conversions. */
3823     if (!is_yuv_fixup(fixup))
3824     {
3825         TRACE("[FAILED]\n");
3826         return FALSE;
3827     }
3828
3829     yuv_fixup = get_yuv_fixup(fixup);
3830     switch(yuv_fixup)
3831     {
3832         case YUV_FIXUP_YUY2:
3833         case YUV_FIXUP_UYVY:
3834         case YUV_FIXUP_YV12:
3835             TRACE("[OK]\n");
3836             return TRUE;
3837
3838         default:
3839             FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3840             TRACE("[FAILED]\n");
3841             return FALSE;
3842     }
3843 }
3844
3845 const struct blit_shader arbfp_blit = {
3846     arbfp_blit_alloc,
3847     arbfp_blit_free,
3848     arbfp_blit_set,
3849     arbfp_blit_unset,
3850     arbfp_blit_color_fixup_supported,
3851 };
3852
3853 #undef GLINFO_LOCATION